linux/mm/percpu.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * mm/percpu.c - percpu memory allocator
   4 *
   5 * Copyright (C) 2009           SUSE Linux Products GmbH
   6 * Copyright (C) 2009           Tejun Heo <tj@kernel.org>
   7 *
   8 * Copyright (C) 2017           Facebook Inc.
   9 * Copyright (C) 2017           Dennis Zhou <dennis@kernel.org>
  10 *
  11 * The percpu allocator handles both static and dynamic areas.  Percpu
  12 * areas are allocated in chunks which are divided into units.  There is
  13 * a 1-to-1 mapping for units to possible cpus.  These units are grouped
  14 * based on NUMA properties of the machine.
  15 *
  16 *  c0                           c1                         c2
  17 *  -------------------          -------------------        ------------
  18 * | u0 | u1 | u2 | u3 |        | u0 | u1 | u2 | u3 |      | u0 | u1 | u
  19 *  -------------------  ......  -------------------  ....  ------------
  20 *
  21 * Allocation is done by offsets into a unit's address space.  Ie., an
  22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
  23 * c1:u1, c1:u2, etc.  On NUMA machines, the mapping may be non-linear
  24 * and even sparse.  Access is handled by configuring percpu base
  25 * registers according to the cpu to unit mappings and offsetting the
  26 * base address using pcpu_unit_size.
  27 *
  28 * There is special consideration for the first chunk which must handle
  29 * the static percpu variables in the kernel image as allocation services
  30 * are not online yet.  In short, the first chunk is structured like so:
  31 *
  32 *                  <Static | [Reserved] | Dynamic>
  33 *
  34 * The static data is copied from the original section managed by the
  35 * linker.  The reserved section, if non-zero, primarily manages static
  36 * percpu variables from kernel modules.  Finally, the dynamic section
  37 * takes care of normal allocations.
  38 *
  39 * The allocator organizes chunks into lists according to free size and
  40 * memcg-awareness.  To make a percpu allocation memcg-aware the __GFP_ACCOUNT
  41 * flag should be passed.  All memcg-aware allocations are sharing one set
  42 * of chunks and all unaccounted allocations and allocations performed
  43 * by processes belonging to the root memory cgroup are using the second set.
  44 *
  45 * The allocator tries to allocate from the fullest chunk first. Each chunk
  46 * is managed by a bitmap with metadata blocks.  The allocation map is updated
  47 * on every allocation and free to reflect the current state while the boundary
  48 * map is only updated on allocation.  Each metadata block contains
  49 * information to help mitigate the need to iterate over large portions
  50 * of the bitmap.  The reverse mapping from page to chunk is stored in
  51 * the page's index.  Lastly, units are lazily backed and grow in unison.
  52 *
  53 * There is a unique conversion that goes on here between bytes and bits.
  54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE.  The chunk
  55 * tracks the number of pages it is responsible for in nr_pages.  Helper
  56 * functions are used to convert from between the bytes, bits, and blocks.
  57 * All hints are managed in bits unless explicitly stated.
  58 *
  59 * To use this allocator, arch code should do the following:
  60 *
  61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
  62 *   regular address to percpu pointer and back if they need to be
  63 *   different from the default
  64 *
  65 * - use pcpu_setup_first_chunk() during percpu area initialization to
  66 *   setup the first chunk containing the kernel static percpu area
  67 */
  68
  69#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  70
  71#include <linux/bitmap.h>
  72#include <linux/cpumask.h>
  73#include <linux/memblock.h>
  74#include <linux/err.h>
  75#include <linux/lcm.h>
  76#include <linux/list.h>
  77#include <linux/log2.h>
  78#include <linux/mm.h>
  79#include <linux/module.h>
  80#include <linux/mutex.h>
  81#include <linux/percpu.h>
  82#include <linux/pfn.h>
  83#include <linux/slab.h>
  84#include <linux/spinlock.h>
  85#include <linux/vmalloc.h>
  86#include <linux/workqueue.h>
  87#include <linux/kmemleak.h>
  88#include <linux/sched.h>
  89#include <linux/sched/mm.h>
  90#include <linux/memcontrol.h>
  91
  92#include <asm/cacheflush.h>
  93#include <asm/sections.h>
  94#include <asm/tlbflush.h>
  95#include <asm/io.h>
  96
  97#define CREATE_TRACE_POINTS
  98#include <trace/events/percpu.h>
  99
 100#include "percpu-internal.h"
 101
 102/*
 103 * The slots are sorted by the size of the biggest continuous free area.
 104 * 1-31 bytes share the same slot.
 105 */
 106#define PCPU_SLOT_BASE_SHIFT            5
 107/* chunks in slots below this are subject to being sidelined on failed alloc */
 108#define PCPU_SLOT_FAIL_THRESHOLD        3
 109
 110#define PCPU_EMPTY_POP_PAGES_LOW        2
 111#define PCPU_EMPTY_POP_PAGES_HIGH       4
 112
 113#ifdef CONFIG_SMP
 114/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
 115#ifndef __addr_to_pcpu_ptr
 116#define __addr_to_pcpu_ptr(addr)                                        \
 117        (void __percpu *)((unsigned long)(addr) -                       \
 118                          (unsigned long)pcpu_base_addr +               \
 119                          (unsigned long)__per_cpu_start)
 120#endif
 121#ifndef __pcpu_ptr_to_addr
 122#define __pcpu_ptr_to_addr(ptr)                                         \
 123        (void __force *)((unsigned long)(ptr) +                         \
 124                         (unsigned long)pcpu_base_addr -                \
 125                         (unsigned long)__per_cpu_start)
 126#endif
 127#else   /* CONFIG_SMP */
 128/* on UP, it's always identity mapped */
 129#define __addr_to_pcpu_ptr(addr)        (void __percpu *)(addr)
 130#define __pcpu_ptr_to_addr(ptr)         (void __force *)(ptr)
 131#endif  /* CONFIG_SMP */
 132
 133static int pcpu_unit_pages __ro_after_init;
 134static int pcpu_unit_size __ro_after_init;
 135static int pcpu_nr_units __ro_after_init;
 136static int pcpu_atom_size __ro_after_init;
 137int pcpu_nr_slots __ro_after_init;
 138static int pcpu_free_slot __ro_after_init;
 139int pcpu_sidelined_slot __ro_after_init;
 140int pcpu_to_depopulate_slot __ro_after_init;
 141static size_t pcpu_chunk_struct_size __ro_after_init;
 142
 143/* cpus with the lowest and highest unit addresses */
 144static unsigned int pcpu_low_unit_cpu __ro_after_init;
 145static unsigned int pcpu_high_unit_cpu __ro_after_init;
 146
 147/* the address of the first chunk which starts with the kernel static area */
 148void *pcpu_base_addr __ro_after_init;
 149EXPORT_SYMBOL_GPL(pcpu_base_addr);
 150
 151static const int *pcpu_unit_map __ro_after_init;                /* cpu -> unit */
 152const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
 153
 154/* group information, used for vm allocation */
 155static int pcpu_nr_groups __ro_after_init;
 156static const unsigned long *pcpu_group_offsets __ro_after_init;
 157static const size_t *pcpu_group_sizes __ro_after_init;
 158
 159/*
 160 * The first chunk which always exists.  Note that unlike other
 161 * chunks, this one can be allocated and mapped in several different
 162 * ways and thus often doesn't live in the vmalloc area.
 163 */
 164struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
 165
 166/*
 167 * Optional reserved chunk.  This chunk reserves part of the first
 168 * chunk and serves it for reserved allocations.  When the reserved
 169 * region doesn't exist, the following variable is NULL.
 170 */
 171struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
 172
 173DEFINE_SPINLOCK(pcpu_lock);     /* all internal data structures */
 174static DEFINE_MUTEX(pcpu_alloc_mutex);  /* chunk create/destroy, [de]pop, map ext */
 175
 176struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
 177
 178/* chunks which need their map areas extended, protected by pcpu_lock */
 179static LIST_HEAD(pcpu_map_extend_chunks);
 180
 181/*
 182 * The number of empty populated pages, protected by pcpu_lock.
 183 * The reserved chunk doesn't contribute to the count.
 184 */
 185int pcpu_nr_empty_pop_pages;
 186
 187/*
 188 * The number of populated pages in use by the allocator, protected by
 189 * pcpu_lock.  This number is kept per a unit per chunk (i.e. when a page gets
 190 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
 191 * and increments/decrements this count by 1).
 192 */
 193static unsigned long pcpu_nr_populated;
 194
 195/*
 196 * Balance work is used to populate or destroy chunks asynchronously.  We
 197 * try to keep the number of populated free pages between
 198 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
 199 * empty chunk.
 200 */
 201static void pcpu_balance_workfn(struct work_struct *work);
 202static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
 203static bool pcpu_async_enabled __read_mostly;
 204static bool pcpu_atomic_alloc_failed;
 205
 206static void pcpu_schedule_balance_work(void)
 207{
 208        if (pcpu_async_enabled)
 209                schedule_work(&pcpu_balance_work);
 210}
 211
 212/**
 213 * pcpu_addr_in_chunk - check if the address is served from this chunk
 214 * @chunk: chunk of interest
 215 * @addr: percpu address
 216 *
 217 * RETURNS:
 218 * True if the address is served from this chunk.
 219 */
 220static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
 221{
 222        void *start_addr, *end_addr;
 223
 224        if (!chunk)
 225                return false;
 226
 227        start_addr = chunk->base_addr + chunk->start_offset;
 228        end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
 229                   chunk->end_offset;
 230
 231        return addr >= start_addr && addr < end_addr;
 232}
 233
 234static int __pcpu_size_to_slot(int size)
 235{
 236        int highbit = fls(size);        /* size is in bytes */
 237        return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
 238}
 239
 240static int pcpu_size_to_slot(int size)
 241{
 242        if (size == pcpu_unit_size)
 243                return pcpu_free_slot;
 244        return __pcpu_size_to_slot(size);
 245}
 246
 247static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
 248{
 249        const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 250
 251        if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
 252            chunk_md->contig_hint == 0)
 253                return 0;
 254
 255        return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
 256}
 257
 258/* set the pointer to a chunk in a page struct */
 259static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
 260{
 261        page->index = (unsigned long)pcpu;
 262}
 263
 264/* obtain pointer to a chunk from a page struct */
 265static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
 266{
 267        return (struct pcpu_chunk *)page->index;
 268}
 269
 270static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
 271{
 272        return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
 273}
 274
 275static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
 276{
 277        return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
 278}
 279
 280static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
 281                                     unsigned int cpu, int page_idx)
 282{
 283        return (unsigned long)chunk->base_addr +
 284               pcpu_unit_page_offset(cpu, page_idx);
 285}
 286
 287/*
 288 * The following are helper functions to help access bitmaps and convert
 289 * between bitmap offsets to address offsets.
 290 */
 291static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
 292{
 293        return chunk->alloc_map +
 294               (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
 295}
 296
 297static unsigned long pcpu_off_to_block_index(int off)
 298{
 299        return off / PCPU_BITMAP_BLOCK_BITS;
 300}
 301
 302static unsigned long pcpu_off_to_block_off(int off)
 303{
 304        return off & (PCPU_BITMAP_BLOCK_BITS - 1);
 305}
 306
 307static unsigned long pcpu_block_off_to_off(int index, int off)
 308{
 309        return index * PCPU_BITMAP_BLOCK_BITS + off;
 310}
 311
 312/**
 313 * pcpu_check_block_hint - check against the contig hint
 314 * @block: block of interest
 315 * @bits: size of allocation
 316 * @align: alignment of area (max PAGE_SIZE)
 317 *
 318 * Check to see if the allocation can fit in the block's contig hint.
 319 * Note, a chunk uses the same hints as a block so this can also check against
 320 * the chunk's contig hint.
 321 */
 322static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
 323                                  size_t align)
 324{
 325        int bit_off = ALIGN(block->contig_hint_start, align) -
 326                block->contig_hint_start;
 327
 328        return bit_off + bits <= block->contig_hint;
 329}
 330
 331/*
 332 * pcpu_next_hint - determine which hint to use
 333 * @block: block of interest
 334 * @alloc_bits: size of allocation
 335 *
 336 * This determines if we should scan based on the scan_hint or first_free.
 337 * In general, we want to scan from first_free to fulfill allocations by
 338 * first fit.  However, if we know a scan_hint at position scan_hint_start
 339 * cannot fulfill an allocation, we can begin scanning from there knowing
 340 * the contig_hint will be our fallback.
 341 */
 342static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
 343{
 344        /*
 345         * The three conditions below determine if we can skip past the
 346         * scan_hint.  First, does the scan hint exist.  Second, is the
 347         * contig_hint after the scan_hint (possibly not true iff
 348         * contig_hint == scan_hint).  Third, is the allocation request
 349         * larger than the scan_hint.
 350         */
 351        if (block->scan_hint &&
 352            block->contig_hint_start > block->scan_hint_start &&
 353            alloc_bits > block->scan_hint)
 354                return block->scan_hint_start + block->scan_hint;
 355
 356        return block->first_free;
 357}
 358
 359/**
 360 * pcpu_next_md_free_region - finds the next hint free area
 361 * @chunk: chunk of interest
 362 * @bit_off: chunk offset
 363 * @bits: size of free area
 364 *
 365 * Helper function for pcpu_for_each_md_free_region.  It checks
 366 * block->contig_hint and performs aggregation across blocks to find the
 367 * next hint.  It modifies bit_off and bits in-place to be consumed in the
 368 * loop.
 369 */
 370static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
 371                                     int *bits)
 372{
 373        int i = pcpu_off_to_block_index(*bit_off);
 374        int block_off = pcpu_off_to_block_off(*bit_off);
 375        struct pcpu_block_md *block;
 376
 377        *bits = 0;
 378        for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 379             block++, i++) {
 380                /* handles contig area across blocks */
 381                if (*bits) {
 382                        *bits += block->left_free;
 383                        if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 384                                continue;
 385                        return;
 386                }
 387
 388                /*
 389                 * This checks three things.  First is there a contig_hint to
 390                 * check.  Second, have we checked this hint before by
 391                 * comparing the block_off.  Third, is this the same as the
 392                 * right contig hint.  In the last case, it spills over into
 393                 * the next block and should be handled by the contig area
 394                 * across blocks code.
 395                 */
 396                *bits = block->contig_hint;
 397                if (*bits && block->contig_hint_start >= block_off &&
 398                    *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
 399                        *bit_off = pcpu_block_off_to_off(i,
 400                                        block->contig_hint_start);
 401                        return;
 402                }
 403                /* reset to satisfy the second predicate above */
 404                block_off = 0;
 405
 406                *bits = block->right_free;
 407                *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
 408        }
 409}
 410
 411/**
 412 * pcpu_next_fit_region - finds fit areas for a given allocation request
 413 * @chunk: chunk of interest
 414 * @alloc_bits: size of allocation
 415 * @align: alignment of area (max PAGE_SIZE)
 416 * @bit_off: chunk offset
 417 * @bits: size of free area
 418 *
 419 * Finds the next free region that is viable for use with a given size and
 420 * alignment.  This only returns if there is a valid area to be used for this
 421 * allocation.  block->first_free is returned if the allocation request fits
 422 * within the block to see if the request can be fulfilled prior to the contig
 423 * hint.
 424 */
 425static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
 426                                 int align, int *bit_off, int *bits)
 427{
 428        int i = pcpu_off_to_block_index(*bit_off);
 429        int block_off = pcpu_off_to_block_off(*bit_off);
 430        struct pcpu_block_md *block;
 431
 432        *bits = 0;
 433        for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 434             block++, i++) {
 435                /* handles contig area across blocks */
 436                if (*bits) {
 437                        *bits += block->left_free;
 438                        if (*bits >= alloc_bits)
 439                                return;
 440                        if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 441                                continue;
 442                }
 443
 444                /* check block->contig_hint */
 445                *bits = ALIGN(block->contig_hint_start, align) -
 446                        block->contig_hint_start;
 447                /*
 448                 * This uses the block offset to determine if this has been
 449                 * checked in the prior iteration.
 450                 */
 451                if (block->contig_hint &&
 452                    block->contig_hint_start >= block_off &&
 453                    block->contig_hint >= *bits + alloc_bits) {
 454                        int start = pcpu_next_hint(block, alloc_bits);
 455
 456                        *bits += alloc_bits + block->contig_hint_start -
 457                                 start;
 458                        *bit_off = pcpu_block_off_to_off(i, start);
 459                        return;
 460                }
 461                /* reset to satisfy the second predicate above */
 462                block_off = 0;
 463
 464                *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
 465                                 align);
 466                *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
 467                *bit_off = pcpu_block_off_to_off(i, *bit_off);
 468                if (*bits >= alloc_bits)
 469                        return;
 470        }
 471
 472        /* no valid offsets were found - fail condition */
 473        *bit_off = pcpu_chunk_map_bits(chunk);
 474}
 475
 476/*
 477 * Metadata free area iterators.  These perform aggregation of free areas
 478 * based on the metadata blocks and return the offset @bit_off and size in
 479 * bits of the free area @bits.  pcpu_for_each_fit_region only returns when
 480 * a fit is found for the allocation request.
 481 */
 482#define pcpu_for_each_md_free_region(chunk, bit_off, bits)              \
 483        for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits));    \
 484             (bit_off) < pcpu_chunk_map_bits((chunk));                  \
 485             (bit_off) += (bits) + 1,                                   \
 486             pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
 487
 488#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits)     \
 489        for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 490                                  &(bits));                                   \
 491             (bit_off) < pcpu_chunk_map_bits((chunk));                        \
 492             (bit_off) += (bits),                                             \
 493             pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 494                                  &(bits)))
 495
 496/**
 497 * pcpu_mem_zalloc - allocate memory
 498 * @size: bytes to allocate
 499 * @gfp: allocation flags
 500 *
 501 * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
 502 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
 503 * This is to facilitate passing through whitelisted flags.  The
 504 * returned memory is always zeroed.
 505 *
 506 * RETURNS:
 507 * Pointer to the allocated area on success, NULL on failure.
 508 */
 509static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
 510{
 511        if (WARN_ON_ONCE(!slab_is_available()))
 512                return NULL;
 513
 514        if (size <= PAGE_SIZE)
 515                return kzalloc(size, gfp);
 516        else
 517                return __vmalloc(size, gfp | __GFP_ZERO);
 518}
 519
 520/**
 521 * pcpu_mem_free - free memory
 522 * @ptr: memory to free
 523 *
 524 * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
 525 */
 526static void pcpu_mem_free(void *ptr)
 527{
 528        kvfree(ptr);
 529}
 530
 531static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
 532                              bool move_front)
 533{
 534        if (chunk != pcpu_reserved_chunk) {
 535                if (move_front)
 536                        list_move(&chunk->list, &pcpu_chunk_lists[slot]);
 537                else
 538                        list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]);
 539        }
 540}
 541
 542static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
 543{
 544        __pcpu_chunk_move(chunk, slot, true);
 545}
 546
 547/**
 548 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
 549 * @chunk: chunk of interest
 550 * @oslot: the previous slot it was on
 551 *
 552 * This function is called after an allocation or free changed @chunk.
 553 * New slot according to the changed state is determined and @chunk is
 554 * moved to the slot.  Note that the reserved chunk is never put on
 555 * chunk slots.
 556 *
 557 * CONTEXT:
 558 * pcpu_lock.
 559 */
 560static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
 561{
 562        int nslot = pcpu_chunk_slot(chunk);
 563
 564        /* leave isolated chunks in-place */
 565        if (chunk->isolated)
 566                return;
 567
 568        if (oslot != nslot)
 569                __pcpu_chunk_move(chunk, nslot, oslot < nslot);
 570}
 571
 572static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
 573{
 574        lockdep_assert_held(&pcpu_lock);
 575
 576        if (!chunk->isolated) {
 577                chunk->isolated = true;
 578                pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
 579        }
 580        list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
 581}
 582
 583static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
 584{
 585        lockdep_assert_held(&pcpu_lock);
 586
 587        if (chunk->isolated) {
 588                chunk->isolated = false;
 589                pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
 590                pcpu_chunk_relocate(chunk, -1);
 591        }
 592}
 593
 594/*
 595 * pcpu_update_empty_pages - update empty page counters
 596 * @chunk: chunk of interest
 597 * @nr: nr of empty pages
 598 *
 599 * This is used to keep track of the empty pages now based on the premise
 600 * a md_block covers a page.  The hint update functions recognize if a block
 601 * is made full or broken to calculate deltas for keeping track of free pages.
 602 */
 603static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
 604{
 605        chunk->nr_empty_pop_pages += nr;
 606        if (chunk != pcpu_reserved_chunk && !chunk->isolated)
 607                pcpu_nr_empty_pop_pages += nr;
 608}
 609
 610/*
 611 * pcpu_region_overlap - determines if two regions overlap
 612 * @a: start of first region, inclusive
 613 * @b: end of first region, exclusive
 614 * @x: start of second region, inclusive
 615 * @y: end of second region, exclusive
 616 *
 617 * This is used to determine if the hint region [a, b) overlaps with the
 618 * allocated region [x, y).
 619 */
 620static inline bool pcpu_region_overlap(int a, int b, int x, int y)
 621{
 622        return (a < y) && (x < b);
 623}
 624
 625/**
 626 * pcpu_block_update - updates a block given a free area
 627 * @block: block of interest
 628 * @start: start offset in block
 629 * @end: end offset in block
 630 *
 631 * Updates a block given a known free area.  The region [start, end) is
 632 * expected to be the entirety of the free area within a block.  Chooses
 633 * the best starting offset if the contig hints are equal.
 634 */
 635static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
 636{
 637        int contig = end - start;
 638
 639        block->first_free = min(block->first_free, start);
 640        if (start == 0)
 641                block->left_free = contig;
 642
 643        if (end == block->nr_bits)
 644                block->right_free = contig;
 645
 646        if (contig > block->contig_hint) {
 647                /* promote the old contig_hint to be the new scan_hint */
 648                if (start > block->contig_hint_start) {
 649                        if (block->contig_hint > block->scan_hint) {
 650                                block->scan_hint_start =
 651                                        block->contig_hint_start;
 652                                block->scan_hint = block->contig_hint;
 653                        } else if (start < block->scan_hint_start) {
 654                                /*
 655                                 * The old contig_hint == scan_hint.  But, the
 656                                 * new contig is larger so hold the invariant
 657                                 * scan_hint_start < contig_hint_start.
 658                                 */
 659                                block->scan_hint = 0;
 660                        }
 661                } else {
 662                        block->scan_hint = 0;
 663                }
 664                block->contig_hint_start = start;
 665                block->contig_hint = contig;
 666        } else if (contig == block->contig_hint) {
 667                if (block->contig_hint_start &&
 668                    (!start ||
 669                     __ffs(start) > __ffs(block->contig_hint_start))) {
 670                        /* start has a better alignment so use it */
 671                        block->contig_hint_start = start;
 672                        if (start < block->scan_hint_start &&
 673                            block->contig_hint > block->scan_hint)
 674                                block->scan_hint = 0;
 675                } else if (start > block->scan_hint_start ||
 676                           block->contig_hint > block->scan_hint) {
 677                        /*
 678                         * Knowing contig == contig_hint, update the scan_hint
 679                         * if it is farther than or larger than the current
 680                         * scan_hint.
 681                         */
 682                        block->scan_hint_start = start;
 683                        block->scan_hint = contig;
 684                }
 685        } else {
 686                /*
 687                 * The region is smaller than the contig_hint.  So only update
 688                 * the scan_hint if it is larger than or equal and farther than
 689                 * the current scan_hint.
 690                 */
 691                if ((start < block->contig_hint_start &&
 692                     (contig > block->scan_hint ||
 693                      (contig == block->scan_hint &&
 694                       start > block->scan_hint_start)))) {
 695                        block->scan_hint_start = start;
 696                        block->scan_hint = contig;
 697                }
 698        }
 699}
 700
 701/*
 702 * pcpu_block_update_scan - update a block given a free area from a scan
 703 * @chunk: chunk of interest
 704 * @bit_off: chunk offset
 705 * @bits: size of free area
 706 *
 707 * Finding the final allocation spot first goes through pcpu_find_block_fit()
 708 * to find a block that can hold the allocation and then pcpu_alloc_area()
 709 * where a scan is used.  When allocations require specific alignments,
 710 * we can inadvertently create holes which will not be seen in the alloc
 711 * or free paths.
 712 *
 713 * This takes a given free area hole and updates a block as it may change the
 714 * scan_hint.  We need to scan backwards to ensure we don't miss free bits
 715 * from alignment.
 716 */
 717static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
 718                                   int bits)
 719{
 720        int s_off = pcpu_off_to_block_off(bit_off);
 721        int e_off = s_off + bits;
 722        int s_index, l_bit;
 723        struct pcpu_block_md *block;
 724
 725        if (e_off > PCPU_BITMAP_BLOCK_BITS)
 726                return;
 727
 728        s_index = pcpu_off_to_block_index(bit_off);
 729        block = chunk->md_blocks + s_index;
 730
 731        /* scan backwards in case of alignment skipping free bits */
 732        l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
 733        s_off = (s_off == l_bit) ? 0 : l_bit + 1;
 734
 735        pcpu_block_update(block, s_off, e_off);
 736}
 737
 738/**
 739 * pcpu_chunk_refresh_hint - updates metadata about a chunk
 740 * @chunk: chunk of interest
 741 * @full_scan: if we should scan from the beginning
 742 *
 743 * Iterates over the metadata blocks to find the largest contig area.
 744 * A full scan can be avoided on the allocation path as this is triggered
 745 * if we broke the contig_hint.  In doing so, the scan_hint will be before
 746 * the contig_hint or after if the scan_hint == contig_hint.  This cannot
 747 * be prevented on freeing as we want to find the largest area possibly
 748 * spanning blocks.
 749 */
 750static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
 751{
 752        struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 753        int bit_off, bits;
 754
 755        /* promote scan_hint to contig_hint */
 756        if (!full_scan && chunk_md->scan_hint) {
 757                bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
 758                chunk_md->contig_hint_start = chunk_md->scan_hint_start;
 759                chunk_md->contig_hint = chunk_md->scan_hint;
 760                chunk_md->scan_hint = 0;
 761        } else {
 762                bit_off = chunk_md->first_free;
 763                chunk_md->contig_hint = 0;
 764        }
 765
 766        bits = 0;
 767        pcpu_for_each_md_free_region(chunk, bit_off, bits)
 768                pcpu_block_update(chunk_md, bit_off, bit_off + bits);
 769}
 770
 771/**
 772 * pcpu_block_refresh_hint
 773 * @chunk: chunk of interest
 774 * @index: index of the metadata block
 775 *
 776 * Scans over the block beginning at first_free and updates the block
 777 * metadata accordingly.
 778 */
 779static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
 780{
 781        struct pcpu_block_md *block = chunk->md_blocks + index;
 782        unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
 783        unsigned int rs, re, start;     /* region start, region end */
 784
 785        /* promote scan_hint to contig_hint */
 786        if (block->scan_hint) {
 787                start = block->scan_hint_start + block->scan_hint;
 788                block->contig_hint_start = block->scan_hint_start;
 789                block->contig_hint = block->scan_hint;
 790                block->scan_hint = 0;
 791        } else {
 792                start = block->first_free;
 793                block->contig_hint = 0;
 794        }
 795
 796        block->right_free = 0;
 797
 798        /* iterate over free areas and update the contig hints */
 799        bitmap_for_each_clear_region(alloc_map, rs, re, start,
 800                                     PCPU_BITMAP_BLOCK_BITS)
 801                pcpu_block_update(block, rs, re);
 802}
 803
 804/**
 805 * pcpu_block_update_hint_alloc - update hint on allocation path
 806 * @chunk: chunk of interest
 807 * @bit_off: chunk offset
 808 * @bits: size of request
 809 *
 810 * Updates metadata for the allocation path.  The metadata only has to be
 811 * refreshed by a full scan iff the chunk's contig hint is broken.  Block level
 812 * scans are required if the block's contig hint is broken.
 813 */
 814static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
 815                                         int bits)
 816{
 817        struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 818        int nr_empty_pages = 0;
 819        struct pcpu_block_md *s_block, *e_block, *block;
 820        int s_index, e_index;   /* block indexes of the freed allocation */
 821        int s_off, e_off;       /* block offsets of the freed allocation */
 822
 823        /*
 824         * Calculate per block offsets.
 825         * The calculation uses an inclusive range, but the resulting offsets
 826         * are [start, end).  e_index always points to the last block in the
 827         * range.
 828         */
 829        s_index = pcpu_off_to_block_index(bit_off);
 830        e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 831        s_off = pcpu_off_to_block_off(bit_off);
 832        e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 833
 834        s_block = chunk->md_blocks + s_index;
 835        e_block = chunk->md_blocks + e_index;
 836
 837        /*
 838         * Update s_block.
 839         * block->first_free must be updated if the allocation takes its place.
 840         * If the allocation breaks the contig_hint, a scan is required to
 841         * restore this hint.
 842         */
 843        if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
 844                nr_empty_pages++;
 845
 846        if (s_off == s_block->first_free)
 847                s_block->first_free = find_next_zero_bit(
 848                                        pcpu_index_alloc_map(chunk, s_index),
 849                                        PCPU_BITMAP_BLOCK_BITS,
 850                                        s_off + bits);
 851
 852        if (pcpu_region_overlap(s_block->scan_hint_start,
 853                                s_block->scan_hint_start + s_block->scan_hint,
 854                                s_off,
 855                                s_off + bits))
 856                s_block->scan_hint = 0;
 857
 858        if (pcpu_region_overlap(s_block->contig_hint_start,
 859                                s_block->contig_hint_start +
 860                                s_block->contig_hint,
 861                                s_off,
 862                                s_off + bits)) {
 863                /* block contig hint is broken - scan to fix it */
 864                if (!s_off)
 865                        s_block->left_free = 0;
 866                pcpu_block_refresh_hint(chunk, s_index);
 867        } else {
 868                /* update left and right contig manually */
 869                s_block->left_free = min(s_block->left_free, s_off);
 870                if (s_index == e_index)
 871                        s_block->right_free = min_t(int, s_block->right_free,
 872                                        PCPU_BITMAP_BLOCK_BITS - e_off);
 873                else
 874                        s_block->right_free = 0;
 875        }
 876
 877        /*
 878         * Update e_block.
 879         */
 880        if (s_index != e_index) {
 881                if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
 882                        nr_empty_pages++;
 883
 884                /*
 885                 * When the allocation is across blocks, the end is along
 886                 * the left part of the e_block.
 887                 */
 888                e_block->first_free = find_next_zero_bit(
 889                                pcpu_index_alloc_map(chunk, e_index),
 890                                PCPU_BITMAP_BLOCK_BITS, e_off);
 891
 892                if (e_off == PCPU_BITMAP_BLOCK_BITS) {
 893                        /* reset the block */
 894                        e_block++;
 895                } else {
 896                        if (e_off > e_block->scan_hint_start)
 897                                e_block->scan_hint = 0;
 898
 899                        e_block->left_free = 0;
 900                        if (e_off > e_block->contig_hint_start) {
 901                                /* contig hint is broken - scan to fix it */
 902                                pcpu_block_refresh_hint(chunk, e_index);
 903                        } else {
 904                                e_block->right_free =
 905                                        min_t(int, e_block->right_free,
 906                                              PCPU_BITMAP_BLOCK_BITS - e_off);
 907                        }
 908                }
 909
 910                /* update in-between md_blocks */
 911                nr_empty_pages += (e_index - s_index - 1);
 912                for (block = s_block + 1; block < e_block; block++) {
 913                        block->scan_hint = 0;
 914                        block->contig_hint = 0;
 915                        block->left_free = 0;
 916                        block->right_free = 0;
 917                }
 918        }
 919
 920        if (nr_empty_pages)
 921                pcpu_update_empty_pages(chunk, -nr_empty_pages);
 922
 923        if (pcpu_region_overlap(chunk_md->scan_hint_start,
 924                                chunk_md->scan_hint_start +
 925                                chunk_md->scan_hint,
 926                                bit_off,
 927                                bit_off + bits))
 928                chunk_md->scan_hint = 0;
 929
 930        /*
 931         * The only time a full chunk scan is required is if the chunk
 932         * contig hint is broken.  Otherwise, it means a smaller space
 933         * was used and therefore the chunk contig hint is still correct.
 934         */
 935        if (pcpu_region_overlap(chunk_md->contig_hint_start,
 936                                chunk_md->contig_hint_start +
 937                                chunk_md->contig_hint,
 938                                bit_off,
 939                                bit_off + bits))
 940                pcpu_chunk_refresh_hint(chunk, false);
 941}
 942
 943/**
 944 * pcpu_block_update_hint_free - updates the block hints on the free path
 945 * @chunk: chunk of interest
 946 * @bit_off: chunk offset
 947 * @bits: size of request
 948 *
 949 * Updates metadata for the allocation path.  This avoids a blind block
 950 * refresh by making use of the block contig hints.  If this fails, it scans
 951 * forward and backward to determine the extent of the free area.  This is
 952 * capped at the boundary of blocks.
 953 *
 954 * A chunk update is triggered if a page becomes free, a block becomes free,
 955 * or the free spans across blocks.  This tradeoff is to minimize iterating
 956 * over the block metadata to update chunk_md->contig_hint.
 957 * chunk_md->contig_hint may be off by up to a page, but it will never be more
 958 * than the available space.  If the contig hint is contained in one block, it
 959 * will be accurate.
 960 */
 961static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
 962                                        int bits)
 963{
 964        int nr_empty_pages = 0;
 965        struct pcpu_block_md *s_block, *e_block, *block;
 966        int s_index, e_index;   /* block indexes of the freed allocation */
 967        int s_off, e_off;       /* block offsets of the freed allocation */
 968        int start, end;         /* start and end of the whole free area */
 969
 970        /*
 971         * Calculate per block offsets.
 972         * The calculation uses an inclusive range, but the resulting offsets
 973         * are [start, end).  e_index always points to the last block in the
 974         * range.
 975         */
 976        s_index = pcpu_off_to_block_index(bit_off);
 977        e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 978        s_off = pcpu_off_to_block_off(bit_off);
 979        e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 980
 981        s_block = chunk->md_blocks + s_index;
 982        e_block = chunk->md_blocks + e_index;
 983
 984        /*
 985         * Check if the freed area aligns with the block->contig_hint.
 986         * If it does, then the scan to find the beginning/end of the
 987         * larger free area can be avoided.
 988         *
 989         * start and end refer to beginning and end of the free area
 990         * within each their respective blocks.  This is not necessarily
 991         * the entire free area as it may span blocks past the beginning
 992         * or end of the block.
 993         */
 994        start = s_off;
 995        if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
 996                start = s_block->contig_hint_start;
 997        } else {
 998                /*
 999                 * Scan backwards to find the extent of the free area.
1000                 * find_last_bit returns the starting bit, so if the start bit
1001                 * is returned, that means there was no last bit and the
1002                 * remainder of the chunk is free.
1003                 */
1004                int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
1005                                          start);
1006                start = (start == l_bit) ? 0 : l_bit + 1;
1007        }
1008
1009        end = e_off;
1010        if (e_off == e_block->contig_hint_start)
1011                end = e_block->contig_hint_start + e_block->contig_hint;
1012        else
1013                end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
1014                                    PCPU_BITMAP_BLOCK_BITS, end);
1015
1016        /* update s_block */
1017        e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1018        if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1019                nr_empty_pages++;
1020        pcpu_block_update(s_block, start, e_off);
1021
1022        /* freeing in the same block */
1023        if (s_index != e_index) {
1024                /* update e_block */
1025                if (end == PCPU_BITMAP_BLOCK_BITS)
1026                        nr_empty_pages++;
1027                pcpu_block_update(e_block, 0, end);
1028
1029                /* reset md_blocks in the middle */
1030                nr_empty_pages += (e_index - s_index - 1);
1031                for (block = s_block + 1; block < e_block; block++) {
1032                        block->first_free = 0;
1033                        block->scan_hint = 0;
1034                        block->contig_hint_start = 0;
1035                        block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1036                        block->left_free = PCPU_BITMAP_BLOCK_BITS;
1037                        block->right_free = PCPU_BITMAP_BLOCK_BITS;
1038                }
1039        }
1040
1041        if (nr_empty_pages)
1042                pcpu_update_empty_pages(chunk, nr_empty_pages);
1043
1044        /*
1045         * Refresh chunk metadata when the free makes a block free or spans
1046         * across blocks.  The contig_hint may be off by up to a page, but if
1047         * the contig_hint is contained in a block, it will be accurate with
1048         * the else condition below.
1049         */
1050        if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1051                pcpu_chunk_refresh_hint(chunk, true);
1052        else
1053                pcpu_block_update(&chunk->chunk_md,
1054                                  pcpu_block_off_to_off(s_index, start),
1055                                  end);
1056}
1057
1058/**
1059 * pcpu_is_populated - determines if the region is populated
1060 * @chunk: chunk of interest
1061 * @bit_off: chunk offset
1062 * @bits: size of area
1063 * @next_off: return value for the next offset to start searching
1064 *
1065 * For atomic allocations, check if the backing pages are populated.
1066 *
1067 * RETURNS:
1068 * Bool if the backing pages are populated.
1069 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1070 */
1071static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1072                              int *next_off)
1073{
1074        unsigned int page_start, page_end, rs, re;
1075
1076        page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1077        page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1078
1079        rs = page_start;
1080        bitmap_next_clear_region(chunk->populated, &rs, &re, page_end);
1081        if (rs >= page_end)
1082                return true;
1083
1084        *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1085        return false;
1086}
1087
1088/**
1089 * pcpu_find_block_fit - finds the block index to start searching
1090 * @chunk: chunk of interest
1091 * @alloc_bits: size of request in allocation units
1092 * @align: alignment of area (max PAGE_SIZE bytes)
1093 * @pop_only: use populated regions only
1094 *
1095 * Given a chunk and an allocation spec, find the offset to begin searching
1096 * for a free region.  This iterates over the bitmap metadata blocks to
1097 * find an offset that will be guaranteed to fit the requirements.  It is
1098 * not quite first fit as if the allocation does not fit in the contig hint
1099 * of a block or chunk, it is skipped.  This errs on the side of caution
1100 * to prevent excess iteration.  Poor alignment can cause the allocator to
1101 * skip over blocks and chunks that have valid free areas.
1102 *
1103 * RETURNS:
1104 * The offset in the bitmap to begin searching.
1105 * -1 if no offset is found.
1106 */
1107static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1108                               size_t align, bool pop_only)
1109{
1110        struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1111        int bit_off, bits, next_off;
1112
1113        /*
1114         * This is an optimization to prevent scanning by assuming if the
1115         * allocation cannot fit in the global hint, there is memory pressure
1116         * and creating a new chunk would happen soon.
1117         */
1118        if (!pcpu_check_block_hint(chunk_md, alloc_bits, align))
1119                return -1;
1120
1121        bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1122        bits = 0;
1123        pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1124                if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1125                                                   &next_off))
1126                        break;
1127
1128                bit_off = next_off;
1129                bits = 0;
1130        }
1131
1132        if (bit_off == pcpu_chunk_map_bits(chunk))
1133                return -1;
1134
1135        return bit_off;
1136}
1137
1138/*
1139 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1140 * @map: the address to base the search on
1141 * @size: the bitmap size in bits
1142 * @start: the bitnumber to start searching at
1143 * @nr: the number of zeroed bits we're looking for
1144 * @align_mask: alignment mask for zero area
1145 * @largest_off: offset of the largest area skipped
1146 * @largest_bits: size of the largest area skipped
1147 *
1148 * The @align_mask should be one less than a power of 2.
1149 *
1150 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1151 * the largest area that was skipped.  This is imperfect, but in general is
1152 * good enough.  The largest remembered region is the largest failed region
1153 * seen.  This does not include anything we possibly skipped due to alignment.
1154 * pcpu_block_update_scan() does scan backwards to try and recover what was
1155 * lost to alignment.  While this can cause scanning to miss earlier possible
1156 * free areas, smaller allocations will eventually fill those holes.
1157 */
1158static unsigned long pcpu_find_zero_area(unsigned long *map,
1159                                         unsigned long size,
1160                                         unsigned long start,
1161                                         unsigned long nr,
1162                                         unsigned long align_mask,
1163                                         unsigned long *largest_off,
1164                                         unsigned long *largest_bits)
1165{
1166        unsigned long index, end, i, area_off, area_bits;
1167again:
1168        index = find_next_zero_bit(map, size, start);
1169
1170        /* Align allocation */
1171        index = __ALIGN_MASK(index, align_mask);
1172        area_off = index;
1173
1174        end = index + nr;
1175        if (end > size)
1176                return end;
1177        i = find_next_bit(map, end, index);
1178        if (i < end) {
1179                area_bits = i - area_off;
1180                /* remember largest unused area with best alignment */
1181                if (area_bits > *largest_bits ||
1182                    (area_bits == *largest_bits && *largest_off &&
1183                     (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1184                        *largest_off = area_off;
1185                        *largest_bits = area_bits;
1186                }
1187
1188                start = i + 1;
1189                goto again;
1190        }
1191        return index;
1192}
1193
1194/**
1195 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1196 * @chunk: chunk of interest
1197 * @alloc_bits: size of request in allocation units
1198 * @align: alignment of area (max PAGE_SIZE)
1199 * @start: bit_off to start searching
1200 *
1201 * This function takes in a @start offset to begin searching to fit an
1202 * allocation of @alloc_bits with alignment @align.  It needs to scan
1203 * the allocation map because if it fits within the block's contig hint,
1204 * @start will be block->first_free. This is an attempt to fill the
1205 * allocation prior to breaking the contig hint.  The allocation and
1206 * boundary maps are updated accordingly if it confirms a valid
1207 * free area.
1208 *
1209 * RETURNS:
1210 * Allocated addr offset in @chunk on success.
1211 * -1 if no matching area is found.
1212 */
1213static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1214                           size_t align, int start)
1215{
1216        struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1217        size_t align_mask = (align) ? (align - 1) : 0;
1218        unsigned long area_off = 0, area_bits = 0;
1219        int bit_off, end, oslot;
1220
1221        lockdep_assert_held(&pcpu_lock);
1222
1223        oslot = pcpu_chunk_slot(chunk);
1224
1225        /*
1226         * Search to find a fit.
1227         */
1228        end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1229                    pcpu_chunk_map_bits(chunk));
1230        bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1231                                      align_mask, &area_off, &area_bits);
1232        if (bit_off >= end)
1233                return -1;
1234
1235        if (area_bits)
1236                pcpu_block_update_scan(chunk, area_off, area_bits);
1237
1238        /* update alloc map */
1239        bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1240
1241        /* update boundary map */
1242        set_bit(bit_off, chunk->bound_map);
1243        bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1244        set_bit(bit_off + alloc_bits, chunk->bound_map);
1245
1246        chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1247
1248        /* update first free bit */
1249        if (bit_off == chunk_md->first_free)
1250                chunk_md->first_free = find_next_zero_bit(
1251                                        chunk->alloc_map,
1252                                        pcpu_chunk_map_bits(chunk),
1253                                        bit_off + alloc_bits);
1254
1255        pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1256
1257        pcpu_chunk_relocate(chunk, oslot);
1258
1259        return bit_off * PCPU_MIN_ALLOC_SIZE;
1260}
1261
1262/**
1263 * pcpu_free_area - frees the corresponding offset
1264 * @chunk: chunk of interest
1265 * @off: addr offset into chunk
1266 *
1267 * This function determines the size of an allocation to free using
1268 * the boundary bitmap and clears the allocation map.
1269 *
1270 * RETURNS:
1271 * Number of freed bytes.
1272 */
1273static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1274{
1275        struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1276        int bit_off, bits, end, oslot, freed;
1277
1278        lockdep_assert_held(&pcpu_lock);
1279        pcpu_stats_area_dealloc(chunk);
1280
1281        oslot = pcpu_chunk_slot(chunk);
1282
1283        bit_off = off / PCPU_MIN_ALLOC_SIZE;
1284
1285        /* find end index */
1286        end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1287                            bit_off + 1);
1288        bits = end - bit_off;
1289        bitmap_clear(chunk->alloc_map, bit_off, bits);
1290
1291        freed = bits * PCPU_MIN_ALLOC_SIZE;
1292
1293        /* update metadata */
1294        chunk->free_bytes += freed;
1295
1296        /* update first free bit */
1297        chunk_md->first_free = min(chunk_md->first_free, bit_off);
1298
1299        pcpu_block_update_hint_free(chunk, bit_off, bits);
1300
1301        pcpu_chunk_relocate(chunk, oslot);
1302
1303        return freed;
1304}
1305
1306static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1307{
1308        block->scan_hint = 0;
1309        block->contig_hint = nr_bits;
1310        block->left_free = nr_bits;
1311        block->right_free = nr_bits;
1312        block->first_free = 0;
1313        block->nr_bits = nr_bits;
1314}
1315
1316static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1317{
1318        struct pcpu_block_md *md_block;
1319
1320        /* init the chunk's block */
1321        pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1322
1323        for (md_block = chunk->md_blocks;
1324             md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1325             md_block++)
1326                pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1327}
1328
1329/**
1330 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1331 * @tmp_addr: the start of the region served
1332 * @map_size: size of the region served
1333 *
1334 * This is responsible for creating the chunks that serve the first chunk.  The
1335 * base_addr is page aligned down of @tmp_addr while the region end is page
1336 * aligned up.  Offsets are kept track of to determine the region served. All
1337 * this is done to appease the bitmap allocator in avoiding partial blocks.
1338 *
1339 * RETURNS:
1340 * Chunk serving the region at @tmp_addr of @map_size.
1341 */
1342static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1343                                                         int map_size)
1344{
1345        struct pcpu_chunk *chunk;
1346        unsigned long aligned_addr, lcm_align;
1347        int start_offset, offset_bits, region_size, region_bits;
1348        size_t alloc_size;
1349
1350        /* region calculations */
1351        aligned_addr = tmp_addr & PAGE_MASK;
1352
1353        start_offset = tmp_addr - aligned_addr;
1354
1355        /*
1356         * Align the end of the region with the LCM of PAGE_SIZE and
1357         * PCPU_BITMAP_BLOCK_SIZE.  One of these constants is a multiple of
1358         * the other.
1359         */
1360        lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1361        region_size = ALIGN(start_offset + map_size, lcm_align);
1362
1363        /* allocate chunk */
1364        alloc_size = struct_size(chunk, populated,
1365                                 BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1366        chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1367        if (!chunk)
1368                panic("%s: Failed to allocate %zu bytes\n", __func__,
1369                      alloc_size);
1370
1371        INIT_LIST_HEAD(&chunk->list);
1372
1373        chunk->base_addr = (void *)aligned_addr;
1374        chunk->start_offset = start_offset;
1375        chunk->end_offset = region_size - chunk->start_offset - map_size;
1376
1377        chunk->nr_pages = region_size >> PAGE_SHIFT;
1378        region_bits = pcpu_chunk_map_bits(chunk);
1379
1380        alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1381        chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1382        if (!chunk->alloc_map)
1383                panic("%s: Failed to allocate %zu bytes\n", __func__,
1384                      alloc_size);
1385
1386        alloc_size =
1387                BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1388        chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1389        if (!chunk->bound_map)
1390                panic("%s: Failed to allocate %zu bytes\n", __func__,
1391                      alloc_size);
1392
1393        alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1394        chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1395        if (!chunk->md_blocks)
1396                panic("%s: Failed to allocate %zu bytes\n", __func__,
1397                      alloc_size);
1398
1399#ifdef CONFIG_MEMCG_KMEM
1400        /* first chunk is free to use */
1401        chunk->obj_cgroups = NULL;
1402#endif
1403        pcpu_init_md_blocks(chunk);
1404
1405        /* manage populated page bitmap */
1406        chunk->immutable = true;
1407        bitmap_fill(chunk->populated, chunk->nr_pages);
1408        chunk->nr_populated = chunk->nr_pages;
1409        chunk->nr_empty_pop_pages = chunk->nr_pages;
1410
1411        chunk->free_bytes = map_size;
1412
1413        if (chunk->start_offset) {
1414                /* hide the beginning of the bitmap */
1415                offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1416                bitmap_set(chunk->alloc_map, 0, offset_bits);
1417                set_bit(0, chunk->bound_map);
1418                set_bit(offset_bits, chunk->bound_map);
1419
1420                chunk->chunk_md.first_free = offset_bits;
1421
1422                pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1423        }
1424
1425        if (chunk->end_offset) {
1426                /* hide the end of the bitmap */
1427                offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1428                bitmap_set(chunk->alloc_map,
1429                           pcpu_chunk_map_bits(chunk) - offset_bits,
1430                           offset_bits);
1431                set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1432                        chunk->bound_map);
1433                set_bit(region_bits, chunk->bound_map);
1434
1435                pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1436                                             - offset_bits, offset_bits);
1437        }
1438
1439        return chunk;
1440}
1441
1442static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1443{
1444        struct pcpu_chunk *chunk;
1445        int region_bits;
1446
1447        chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1448        if (!chunk)
1449                return NULL;
1450
1451        INIT_LIST_HEAD(&chunk->list);
1452        chunk->nr_pages = pcpu_unit_pages;
1453        region_bits = pcpu_chunk_map_bits(chunk);
1454
1455        chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1456                                           sizeof(chunk->alloc_map[0]), gfp);
1457        if (!chunk->alloc_map)
1458                goto alloc_map_fail;
1459
1460        chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1461                                           sizeof(chunk->bound_map[0]), gfp);
1462        if (!chunk->bound_map)
1463                goto bound_map_fail;
1464
1465        chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1466                                           sizeof(chunk->md_blocks[0]), gfp);
1467        if (!chunk->md_blocks)
1468                goto md_blocks_fail;
1469
1470#ifdef CONFIG_MEMCG_KMEM
1471        if (!mem_cgroup_kmem_disabled()) {
1472                chunk->obj_cgroups =
1473                        pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1474                                        sizeof(struct obj_cgroup *), gfp);
1475                if (!chunk->obj_cgroups)
1476                        goto objcg_fail;
1477        }
1478#endif
1479
1480        pcpu_init_md_blocks(chunk);
1481
1482        /* init metadata */
1483        chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1484
1485        return chunk;
1486
1487#ifdef CONFIG_MEMCG_KMEM
1488objcg_fail:
1489        pcpu_mem_free(chunk->md_blocks);
1490#endif
1491md_blocks_fail:
1492        pcpu_mem_free(chunk->bound_map);
1493bound_map_fail:
1494        pcpu_mem_free(chunk->alloc_map);
1495alloc_map_fail:
1496        pcpu_mem_free(chunk);
1497
1498        return NULL;
1499}
1500
1501static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1502{
1503        if (!chunk)
1504                return;
1505#ifdef CONFIG_MEMCG_KMEM
1506        pcpu_mem_free(chunk->obj_cgroups);
1507#endif
1508        pcpu_mem_free(chunk->md_blocks);
1509        pcpu_mem_free(chunk->bound_map);
1510        pcpu_mem_free(chunk->alloc_map);
1511        pcpu_mem_free(chunk);
1512}
1513
1514/**
1515 * pcpu_chunk_populated - post-population bookkeeping
1516 * @chunk: pcpu_chunk which got populated
1517 * @page_start: the start page
1518 * @page_end: the end page
1519 *
1520 * Pages in [@page_start,@page_end) have been populated to @chunk.  Update
1521 * the bookkeeping information accordingly.  Must be called after each
1522 * successful population.
1523 *
1524 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1525 * is to serve an allocation in that area.
1526 */
1527static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1528                                 int page_end)
1529{
1530        int nr = page_end - page_start;
1531
1532        lockdep_assert_held(&pcpu_lock);
1533
1534        bitmap_set(chunk->populated, page_start, nr);
1535        chunk->nr_populated += nr;
1536        pcpu_nr_populated += nr;
1537
1538        pcpu_update_empty_pages(chunk, nr);
1539}
1540
1541/**
1542 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1543 * @chunk: pcpu_chunk which got depopulated
1544 * @page_start: the start page
1545 * @page_end: the end page
1546 *
1547 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1548 * Update the bookkeeping information accordingly.  Must be called after
1549 * each successful depopulation.
1550 */
1551static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1552                                   int page_start, int page_end)
1553{
1554        int nr = page_end - page_start;
1555
1556        lockdep_assert_held(&pcpu_lock);
1557
1558        bitmap_clear(chunk->populated, page_start, nr);
1559        chunk->nr_populated -= nr;
1560        pcpu_nr_populated -= nr;
1561
1562        pcpu_update_empty_pages(chunk, -nr);
1563}
1564
1565/*
1566 * Chunk management implementation.
1567 *
1568 * To allow different implementations, chunk alloc/free and
1569 * [de]population are implemented in a separate file which is pulled
1570 * into this file and compiled together.  The following functions
1571 * should be implemented.
1572 *
1573 * pcpu_populate_chunk          - populate the specified range of a chunk
1574 * pcpu_depopulate_chunk        - depopulate the specified range of a chunk
1575 * pcpu_post_unmap_tlb_flush    - flush tlb for the specified range of a chunk
1576 * pcpu_create_chunk            - create a new chunk
1577 * pcpu_destroy_chunk           - destroy a chunk, always preceded by full depop
1578 * pcpu_addr_to_page            - translate address to physical address
1579 * pcpu_verify_alloc_info       - check alloc_info is acceptable during init
1580 */
1581static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1582                               int page_start, int page_end, gfp_t gfp);
1583static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1584                                  int page_start, int page_end);
1585static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1586                                      int page_start, int page_end);
1587static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1588static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1589static struct page *pcpu_addr_to_page(void *addr);
1590static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1591
1592#ifdef CONFIG_NEED_PER_CPU_KM
1593#include "percpu-km.c"
1594#else
1595#include "percpu-vm.c"
1596#endif
1597
1598/**
1599 * pcpu_chunk_addr_search - determine chunk containing specified address
1600 * @addr: address for which the chunk needs to be determined.
1601 *
1602 * This is an internal function that handles all but static allocations.
1603 * Static percpu address values should never be passed into the allocator.
1604 *
1605 * RETURNS:
1606 * The address of the found chunk.
1607 */
1608static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1609{
1610        /* is it in the dynamic region (first chunk)? */
1611        if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1612                return pcpu_first_chunk;
1613
1614        /* is it in the reserved region? */
1615        if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1616                return pcpu_reserved_chunk;
1617
1618        /*
1619         * The address is relative to unit0 which might be unused and
1620         * thus unmapped.  Offset the address to the unit space of the
1621         * current processor before looking it up in the vmalloc
1622         * space.  Note that any possible cpu id can be used here, so
1623         * there's no need to worry about preemption or cpu hotplug.
1624         */
1625        addr += pcpu_unit_offsets[raw_smp_processor_id()];
1626        return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1627}
1628
1629#ifdef CONFIG_MEMCG_KMEM
1630static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1631                                      struct obj_cgroup **objcgp)
1632{
1633        struct obj_cgroup *objcg;
1634
1635        if (!memcg_kmem_enabled() || !(gfp & __GFP_ACCOUNT))
1636                return true;
1637
1638        objcg = get_obj_cgroup_from_current();
1639        if (!objcg)
1640                return true;
1641
1642        if (obj_cgroup_charge(objcg, gfp, size * num_possible_cpus())) {
1643                obj_cgroup_put(objcg);
1644                return false;
1645        }
1646
1647        *objcgp = objcg;
1648        return true;
1649}
1650
1651static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1652                                       struct pcpu_chunk *chunk, int off,
1653                                       size_t size)
1654{
1655        if (!objcg)
1656                return;
1657
1658        if (likely(chunk && chunk->obj_cgroups)) {
1659                chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1660
1661                rcu_read_lock();
1662                mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1663                                size * num_possible_cpus());
1664                rcu_read_unlock();
1665        } else {
1666                obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1667                obj_cgroup_put(objcg);
1668        }
1669}
1670
1671static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1672{
1673        struct obj_cgroup *objcg;
1674
1675        if (unlikely(!chunk->obj_cgroups))
1676                return;
1677
1678        objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1679        if (!objcg)
1680                return;
1681        chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1682
1683        obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1684
1685        rcu_read_lock();
1686        mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1687                        -(size * num_possible_cpus()));
1688        rcu_read_unlock();
1689
1690        obj_cgroup_put(objcg);
1691}
1692
1693#else /* CONFIG_MEMCG_KMEM */
1694static bool
1695pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1696{
1697        return true;
1698}
1699
1700static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1701                                       struct pcpu_chunk *chunk, int off,
1702                                       size_t size)
1703{
1704}
1705
1706static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1707{
1708}
1709#endif /* CONFIG_MEMCG_KMEM */
1710
1711/**
1712 * pcpu_alloc - the percpu allocator
1713 * @size: size of area to allocate in bytes
1714 * @align: alignment of area (max PAGE_SIZE)
1715 * @reserved: allocate from the reserved chunk if available
1716 * @gfp: allocation flags
1717 *
1718 * Allocate percpu area of @size bytes aligned at @align.  If @gfp doesn't
1719 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1720 * then no warning will be triggered on invalid or failed allocation
1721 * requests.
1722 *
1723 * RETURNS:
1724 * Percpu pointer to the allocated area on success, NULL on failure.
1725 */
1726static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1727                                 gfp_t gfp)
1728{
1729        gfp_t pcpu_gfp;
1730        bool is_atomic;
1731        bool do_warn;
1732        struct obj_cgroup *objcg = NULL;
1733        static int warn_limit = 10;
1734        struct pcpu_chunk *chunk, *next;
1735        const char *err;
1736        int slot, off, cpu, ret;
1737        unsigned long flags;
1738        void __percpu *ptr;
1739        size_t bits, bit_align;
1740
1741        gfp = current_gfp_context(gfp);
1742        /* whitelisted flags that can be passed to the backing allocators */
1743        pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1744        is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1745        do_warn = !(gfp & __GFP_NOWARN);
1746
1747        /*
1748         * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1749         * therefore alignment must be a minimum of that many bytes.
1750         * An allocation may have internal fragmentation from rounding up
1751         * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1752         */
1753        if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1754                align = PCPU_MIN_ALLOC_SIZE;
1755
1756        size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1757        bits = size >> PCPU_MIN_ALLOC_SHIFT;
1758        bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1759
1760        if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1761                     !is_power_of_2(align))) {
1762                WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1763                     size, align);
1764                return NULL;
1765        }
1766
1767        if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1768                return NULL;
1769
1770        if (!is_atomic) {
1771                /*
1772                 * pcpu_balance_workfn() allocates memory under this mutex,
1773                 * and it may wait for memory reclaim. Allow current task
1774                 * to become OOM victim, in case of memory pressure.
1775                 */
1776                if (gfp & __GFP_NOFAIL) {
1777                        mutex_lock(&pcpu_alloc_mutex);
1778                } else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1779                        pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1780                        return NULL;
1781                }
1782        }
1783
1784        spin_lock_irqsave(&pcpu_lock, flags);
1785
1786        /* serve reserved allocations from the reserved chunk if available */
1787        if (reserved && pcpu_reserved_chunk) {
1788                chunk = pcpu_reserved_chunk;
1789
1790                off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1791                if (off < 0) {
1792                        err = "alloc from reserved chunk failed";
1793                        goto fail_unlock;
1794                }
1795
1796                off = pcpu_alloc_area(chunk, bits, bit_align, off);
1797                if (off >= 0)
1798                        goto area_found;
1799
1800                err = "alloc from reserved chunk failed";
1801                goto fail_unlock;
1802        }
1803
1804restart:
1805        /* search through normal chunks */
1806        for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1807                list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1808                                         list) {
1809                        off = pcpu_find_block_fit(chunk, bits, bit_align,
1810                                                  is_atomic);
1811                        if (off < 0) {
1812                                if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1813                                        pcpu_chunk_move(chunk, 0);
1814                                continue;
1815                        }
1816
1817                        off = pcpu_alloc_area(chunk, bits, bit_align, off);
1818                        if (off >= 0) {
1819                                pcpu_reintegrate_chunk(chunk);
1820                                goto area_found;
1821                        }
1822                }
1823        }
1824
1825        spin_unlock_irqrestore(&pcpu_lock, flags);
1826
1827        /*
1828         * No space left.  Create a new chunk.  We don't want multiple
1829         * tasks to create chunks simultaneously.  Serialize and create iff
1830         * there's still no empty chunk after grabbing the mutex.
1831         */
1832        if (is_atomic) {
1833                err = "atomic alloc failed, no space left";
1834                goto fail;
1835        }
1836
1837        if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) {
1838                chunk = pcpu_create_chunk(pcpu_gfp);
1839                if (!chunk) {
1840                        err = "failed to allocate new chunk";
1841                        goto fail;
1842                }
1843
1844                spin_lock_irqsave(&pcpu_lock, flags);
1845                pcpu_chunk_relocate(chunk, -1);
1846        } else {
1847                spin_lock_irqsave(&pcpu_lock, flags);
1848        }
1849
1850        goto restart;
1851
1852area_found:
1853        pcpu_stats_area_alloc(chunk, size);
1854        spin_unlock_irqrestore(&pcpu_lock, flags);
1855
1856        /* populate if not all pages are already there */
1857        if (!is_atomic) {
1858                unsigned int page_start, page_end, rs, re;
1859
1860                page_start = PFN_DOWN(off);
1861                page_end = PFN_UP(off + size);
1862
1863                bitmap_for_each_clear_region(chunk->populated, rs, re,
1864                                             page_start, page_end) {
1865                        WARN_ON(chunk->immutable);
1866
1867                        ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1868
1869                        spin_lock_irqsave(&pcpu_lock, flags);
1870                        if (ret) {
1871                                pcpu_free_area(chunk, off);
1872                                err = "failed to populate";
1873                                goto fail_unlock;
1874                        }
1875                        pcpu_chunk_populated(chunk, rs, re);
1876                        spin_unlock_irqrestore(&pcpu_lock, flags);
1877                }
1878
1879                mutex_unlock(&pcpu_alloc_mutex);
1880        }
1881
1882        if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1883                pcpu_schedule_balance_work();
1884
1885        /* clear the areas and return address relative to base address */
1886        for_each_possible_cpu(cpu)
1887                memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1888
1889        ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1890        kmemleak_alloc_percpu(ptr, size, gfp);
1891
1892        trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1893                        chunk->base_addr, off, ptr);
1894
1895        pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1896
1897        return ptr;
1898
1899fail_unlock:
1900        spin_unlock_irqrestore(&pcpu_lock, flags);
1901fail:
1902        trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1903
1904        if (!is_atomic && do_warn && warn_limit) {
1905                pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1906                        size, align, is_atomic, err);
1907                dump_stack();
1908                if (!--warn_limit)
1909                        pr_info("limit reached, disable warning\n");
1910        }
1911        if (is_atomic) {
1912                /* see the flag handling in pcpu_balance_workfn() */
1913                pcpu_atomic_alloc_failed = true;
1914                pcpu_schedule_balance_work();
1915        } else {
1916                mutex_unlock(&pcpu_alloc_mutex);
1917        }
1918
1919        pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1920
1921        return NULL;
1922}
1923
1924/**
1925 * __alloc_percpu_gfp - allocate dynamic percpu area
1926 * @size: size of area to allocate in bytes
1927 * @align: alignment of area (max PAGE_SIZE)
1928 * @gfp: allocation flags
1929 *
1930 * Allocate zero-filled percpu area of @size bytes aligned at @align.  If
1931 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1932 * be called from any context but is a lot more likely to fail. If @gfp
1933 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1934 * allocation requests.
1935 *
1936 * RETURNS:
1937 * Percpu pointer to the allocated area on success, NULL on failure.
1938 */
1939void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1940{
1941        return pcpu_alloc(size, align, false, gfp);
1942}
1943EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1944
1945/**
1946 * __alloc_percpu - allocate dynamic percpu area
1947 * @size: size of area to allocate in bytes
1948 * @align: alignment of area (max PAGE_SIZE)
1949 *
1950 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1951 */
1952void __percpu *__alloc_percpu(size_t size, size_t align)
1953{
1954        return pcpu_alloc(size, align, false, GFP_KERNEL);
1955}
1956EXPORT_SYMBOL_GPL(__alloc_percpu);
1957
1958/**
1959 * __alloc_reserved_percpu - allocate reserved percpu area
1960 * @size: size of area to allocate in bytes
1961 * @align: alignment of area (max PAGE_SIZE)
1962 *
1963 * Allocate zero-filled percpu area of @size bytes aligned at @align
1964 * from reserved percpu area if arch has set it up; otherwise,
1965 * allocation is served from the same dynamic area.  Might sleep.
1966 * Might trigger writeouts.
1967 *
1968 * CONTEXT:
1969 * Does GFP_KERNEL allocation.
1970 *
1971 * RETURNS:
1972 * Percpu pointer to the allocated area on success, NULL on failure.
1973 */
1974void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1975{
1976        return pcpu_alloc(size, align, true, GFP_KERNEL);
1977}
1978
1979/**
1980 * pcpu_balance_free - manage the amount of free chunks
1981 * @empty_only: free chunks only if there are no populated pages
1982 *
1983 * If empty_only is %false, reclaim all fully free chunks regardless of the
1984 * number of populated pages.  Otherwise, only reclaim chunks that have no
1985 * populated pages.
1986 *
1987 * CONTEXT:
1988 * pcpu_lock (can be dropped temporarily)
1989 */
1990static void pcpu_balance_free(bool empty_only)
1991{
1992        LIST_HEAD(to_free);
1993        struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1994        struct pcpu_chunk *chunk, *next;
1995
1996        lockdep_assert_held(&pcpu_lock);
1997
1998        /*
1999         * There's no reason to keep around multiple unused chunks and VM
2000         * areas can be scarce.  Destroy all free chunks except for one.
2001         */
2002        list_for_each_entry_safe(chunk, next, free_head, list) {
2003                WARN_ON(chunk->immutable);
2004
2005                /* spare the first one */
2006                if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
2007                        continue;
2008
2009                if (!empty_only || chunk->nr_empty_pop_pages == 0)
2010                        list_move(&chunk->list, &to_free);
2011        }
2012
2013        if (list_empty(&to_free))
2014                return;
2015
2016        spin_unlock_irq(&pcpu_lock);
2017        list_for_each_entry_safe(chunk, next, &to_free, list) {
2018                unsigned int rs, re;
2019
2020                bitmap_for_each_set_region(chunk->populated, rs, re, 0,
2021                                           chunk->nr_pages) {
2022                        pcpu_depopulate_chunk(chunk, rs, re);
2023                        spin_lock_irq(&pcpu_lock);
2024                        pcpu_chunk_depopulated(chunk, rs, re);
2025                        spin_unlock_irq(&pcpu_lock);
2026                }
2027                pcpu_destroy_chunk(chunk);
2028                cond_resched();
2029        }
2030        spin_lock_irq(&pcpu_lock);
2031}
2032
2033/**
2034 * pcpu_balance_populated - manage the amount of populated pages
2035 *
2036 * Maintain a certain amount of populated pages to satisfy atomic allocations.
2037 * It is possible that this is called when physical memory is scarce causing
2038 * OOM killer to be triggered.  We should avoid doing so until an actual
2039 * allocation causes the failure as it is possible that requests can be
2040 * serviced from already backed regions.
2041 *
2042 * CONTEXT:
2043 * pcpu_lock (can be dropped temporarily)
2044 */
2045static void pcpu_balance_populated(void)
2046{
2047        /* gfp flags passed to underlying allocators */
2048        const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2049        struct pcpu_chunk *chunk;
2050        int slot, nr_to_pop, ret;
2051
2052        lockdep_assert_held(&pcpu_lock);
2053
2054        /*
2055         * Ensure there are certain number of free populated pages for
2056         * atomic allocs.  Fill up from the most packed so that atomic
2057         * allocs don't increase fragmentation.  If atomic allocation
2058         * failed previously, always populate the maximum amount.  This
2059         * should prevent atomic allocs larger than PAGE_SIZE from keeping
2060         * failing indefinitely; however, large atomic allocs are not
2061         * something we support properly and can be highly unreliable and
2062         * inefficient.
2063         */
2064retry_pop:
2065        if (pcpu_atomic_alloc_failed) {
2066                nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2067                /* best effort anyway, don't worry about synchronization */
2068                pcpu_atomic_alloc_failed = false;
2069        } else {
2070                nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2071                                  pcpu_nr_empty_pop_pages,
2072                                  0, PCPU_EMPTY_POP_PAGES_HIGH);
2073        }
2074
2075        for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2076                unsigned int nr_unpop = 0, rs, re;
2077
2078                if (!nr_to_pop)
2079                        break;
2080
2081                list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
2082                        nr_unpop = chunk->nr_pages - chunk->nr_populated;
2083                        if (nr_unpop)
2084                                break;
2085                }
2086
2087                if (!nr_unpop)
2088                        continue;
2089
2090                /* @chunk can't go away while pcpu_alloc_mutex is held */
2091                bitmap_for_each_clear_region(chunk->populated, rs, re, 0,
2092                                             chunk->nr_pages) {
2093                        int nr = min_t(int, re - rs, nr_to_pop);
2094
2095                        spin_unlock_irq(&pcpu_lock);
2096                        ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2097                        cond_resched();
2098                        spin_lock_irq(&pcpu_lock);
2099                        if (!ret) {
2100                                nr_to_pop -= nr;
2101                                pcpu_chunk_populated(chunk, rs, rs + nr);
2102                        } else {
2103                                nr_to_pop = 0;
2104                        }
2105
2106                        if (!nr_to_pop)
2107                                break;
2108                }
2109        }
2110
2111        if (nr_to_pop) {
2112                /* ran out of chunks to populate, create a new one and retry */
2113                spin_unlock_irq(&pcpu_lock);
2114                chunk = pcpu_create_chunk(gfp);
2115                cond_resched();
2116                spin_lock_irq(&pcpu_lock);
2117                if (chunk) {
2118                        pcpu_chunk_relocate(chunk, -1);
2119                        goto retry_pop;
2120                }
2121        }
2122}
2123
2124/**
2125 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2126 *
2127 * Scan over chunks in the depopulate list and try to release unused populated
2128 * pages back to the system.  Depopulated chunks are sidelined to prevent
2129 * repopulating these pages unless required.  Fully free chunks are reintegrated
2130 * and freed accordingly (1 is kept around).  If we drop below the empty
2131 * populated pages threshold, reintegrate the chunk if it has empty free pages.
2132 * Each chunk is scanned in the reverse order to keep populated pages close to
2133 * the beginning of the chunk.
2134 *
2135 * CONTEXT:
2136 * pcpu_lock (can be dropped temporarily)
2137 *
2138 */
2139static void pcpu_reclaim_populated(void)
2140{
2141        struct pcpu_chunk *chunk;
2142        struct pcpu_block_md *block;
2143        int freed_page_start, freed_page_end;
2144        int i, end;
2145        bool reintegrate;
2146
2147        lockdep_assert_held(&pcpu_lock);
2148
2149        /*
2150         * Once a chunk is isolated to the to_depopulate list, the chunk is no
2151         * longer discoverable to allocations whom may populate pages.  The only
2152         * other accessor is the free path which only returns area back to the
2153         * allocator not touching the populated bitmap.
2154         */
2155        while (!list_empty(&pcpu_chunk_lists[pcpu_to_depopulate_slot])) {
2156                chunk = list_first_entry(&pcpu_chunk_lists[pcpu_to_depopulate_slot],
2157                                         struct pcpu_chunk, list);
2158                WARN_ON(chunk->immutable);
2159
2160                /*
2161                 * Scan chunk's pages in the reverse order to keep populated
2162                 * pages close to the beginning of the chunk.
2163                 */
2164                freed_page_start = chunk->nr_pages;
2165                freed_page_end = 0;
2166                reintegrate = false;
2167                for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2168                        /* no more work to do */
2169                        if (chunk->nr_empty_pop_pages == 0)
2170                                break;
2171
2172                        /* reintegrate chunk to prevent atomic alloc failures */
2173                        if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2174                                reintegrate = true;
2175                                goto end_chunk;
2176                        }
2177
2178                        /*
2179                         * If the page is empty and populated, start or
2180                         * extend the (i, end) range.  If i == 0, decrease
2181                         * i and perform the depopulation to cover the last
2182                         * (first) page in the chunk.
2183                         */
2184                        block = chunk->md_blocks + i;
2185                        if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2186                            test_bit(i, chunk->populated)) {
2187                                if (end == -1)
2188                                        end = i;
2189                                if (i > 0)
2190                                        continue;
2191                                i--;
2192                        }
2193
2194                        /* depopulate if there is an active range */
2195                        if (end == -1)
2196                                continue;
2197
2198                        spin_unlock_irq(&pcpu_lock);
2199                        pcpu_depopulate_chunk(chunk, i + 1, end + 1);
2200                        cond_resched();
2201                        spin_lock_irq(&pcpu_lock);
2202
2203                        pcpu_chunk_depopulated(chunk, i + 1, end + 1);
2204                        freed_page_start = min(freed_page_start, i + 1);
2205                        freed_page_end = max(freed_page_end, end + 1);
2206
2207                        /* reset the range and continue */
2208                        end = -1;
2209                }
2210
2211end_chunk:
2212                /* batch tlb flush per chunk to amortize cost */
2213                if (freed_page_start < freed_page_end) {
2214                        spin_unlock_irq(&pcpu_lock);
2215                        pcpu_post_unmap_tlb_flush(chunk,
2216                                                  freed_page_start,
2217                                                  freed_page_end);
2218                        cond_resched();
2219                        spin_lock_irq(&pcpu_lock);
2220                }
2221
2222                if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2223                        pcpu_reintegrate_chunk(chunk);
2224                else
2225                        list_move_tail(&chunk->list,
2226                                       &pcpu_chunk_lists[pcpu_sidelined_slot]);
2227        }
2228}
2229
2230/**
2231 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2232 * @work: unused
2233 *
2234 * For each chunk type, manage the number of fully free chunks and the number of
2235 * populated pages.  An important thing to consider is when pages are freed and
2236 * how they contribute to the global counts.
2237 */
2238static void pcpu_balance_workfn(struct work_struct *work)
2239{
2240        /*
2241         * pcpu_balance_free() is called twice because the first time we may
2242         * trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2243         * to grow other chunks.  This then gives pcpu_reclaim_populated() time
2244         * to move fully free chunks to the active list to be freed if
2245         * appropriate.
2246         */
2247        mutex_lock(&pcpu_alloc_mutex);
2248        spin_lock_irq(&pcpu_lock);
2249
2250        pcpu_balance_free(false);
2251        pcpu_reclaim_populated();
2252        pcpu_balance_populated();
2253        pcpu_balance_free(true);
2254
2255        spin_unlock_irq(&pcpu_lock);
2256        mutex_unlock(&pcpu_alloc_mutex);
2257}
2258
2259/**
2260 * free_percpu - free percpu area
2261 * @ptr: pointer to area to free
2262 *
2263 * Free percpu area @ptr.
2264 *
2265 * CONTEXT:
2266 * Can be called from atomic context.
2267 */
2268void free_percpu(void __percpu *ptr)
2269{
2270        void *addr;
2271        struct pcpu_chunk *chunk;
2272        unsigned long flags;
2273        int size, off;
2274        bool need_balance = false;
2275
2276        if (!ptr)
2277                return;
2278
2279        kmemleak_free_percpu(ptr);
2280
2281        addr = __pcpu_ptr_to_addr(ptr);
2282
2283        spin_lock_irqsave(&pcpu_lock, flags);
2284
2285        chunk = pcpu_chunk_addr_search(addr);
2286        off = addr - chunk->base_addr;
2287
2288        size = pcpu_free_area(chunk, off);
2289
2290        pcpu_memcg_free_hook(chunk, off, size);
2291
2292        /*
2293         * If there are more than one fully free chunks, wake up grim reaper.
2294         * If the chunk is isolated, it may be in the process of being
2295         * reclaimed.  Let reclaim manage cleaning up of that chunk.
2296         */
2297        if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2298                struct pcpu_chunk *pos;
2299
2300                list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2301                        if (pos != chunk) {
2302                                need_balance = true;
2303                                break;
2304                        }
2305        } else if (pcpu_should_reclaim_chunk(chunk)) {
2306                pcpu_isolate_chunk(chunk);
2307                need_balance = true;
2308        }
2309
2310        trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2311
2312        spin_unlock_irqrestore(&pcpu_lock, flags);
2313
2314        if (need_balance)
2315                pcpu_schedule_balance_work();
2316}
2317EXPORT_SYMBOL_GPL(free_percpu);
2318
2319bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2320{
2321#ifdef CONFIG_SMP
2322        const size_t static_size = __per_cpu_end - __per_cpu_start;
2323        void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2324        unsigned int cpu;
2325
2326        for_each_possible_cpu(cpu) {
2327                void *start = per_cpu_ptr(base, cpu);
2328                void *va = (void *)addr;
2329
2330                if (va >= start && va < start + static_size) {
2331                        if (can_addr) {
2332                                *can_addr = (unsigned long) (va - start);
2333                                *can_addr += (unsigned long)
2334                                        per_cpu_ptr(base, get_boot_cpu_id());
2335                        }
2336                        return true;
2337                }
2338        }
2339#endif
2340        /* on UP, can't distinguish from other static vars, always false */
2341        return false;
2342}
2343
2344/**
2345 * is_kernel_percpu_address - test whether address is from static percpu area
2346 * @addr: address to test
2347 *
2348 * Test whether @addr belongs to in-kernel static percpu area.  Module
2349 * static percpu areas are not considered.  For those, use
2350 * is_module_percpu_address().
2351 *
2352 * RETURNS:
2353 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2354 */
2355bool is_kernel_percpu_address(unsigned long addr)
2356{
2357        return __is_kernel_percpu_address(addr, NULL);
2358}
2359
2360/**
2361 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2362 * @addr: the address to be converted to physical address
2363 *
2364 * Given @addr which is dereferenceable address obtained via one of
2365 * percpu access macros, this function translates it into its physical
2366 * address.  The caller is responsible for ensuring @addr stays valid
2367 * until this function finishes.
2368 *
2369 * percpu allocator has special setup for the first chunk, which currently
2370 * supports either embedding in linear address space or vmalloc mapping,
2371 * and, from the second one, the backing allocator (currently either vm or
2372 * km) provides translation.
2373 *
2374 * The addr can be translated simply without checking if it falls into the
2375 * first chunk. But the current code reflects better how percpu allocator
2376 * actually works, and the verification can discover both bugs in percpu
2377 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2378 * code.
2379 *
2380 * RETURNS:
2381 * The physical address for @addr.
2382 */
2383phys_addr_t per_cpu_ptr_to_phys(void *addr)
2384{
2385        void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2386        bool in_first_chunk = false;
2387        unsigned long first_low, first_high;
2388        unsigned int cpu;
2389
2390        /*
2391         * The following test on unit_low/high isn't strictly
2392         * necessary but will speed up lookups of addresses which
2393         * aren't in the first chunk.
2394         *
2395         * The address check is against full chunk sizes.  pcpu_base_addr
2396         * points to the beginning of the first chunk including the
2397         * static region.  Assumes good intent as the first chunk may
2398         * not be full (ie. < pcpu_unit_pages in size).
2399         */
2400        first_low = (unsigned long)pcpu_base_addr +
2401                    pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2402        first_high = (unsigned long)pcpu_base_addr +
2403                     pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2404        if ((unsigned long)addr >= first_low &&
2405            (unsigned long)addr < first_high) {
2406                for_each_possible_cpu(cpu) {
2407                        void *start = per_cpu_ptr(base, cpu);
2408
2409                        if (addr >= start && addr < start + pcpu_unit_size) {
2410                                in_first_chunk = true;
2411                                break;
2412                        }
2413                }
2414        }
2415
2416        if (in_first_chunk) {
2417                if (!is_vmalloc_addr(addr))
2418                        return __pa(addr);
2419                else
2420                        return page_to_phys(vmalloc_to_page(addr)) +
2421                               offset_in_page(addr);
2422        } else
2423                return page_to_phys(pcpu_addr_to_page(addr)) +
2424                       offset_in_page(addr);
2425}
2426
2427/**
2428 * pcpu_alloc_alloc_info - allocate percpu allocation info
2429 * @nr_groups: the number of groups
2430 * @nr_units: the number of units
2431 *
2432 * Allocate ai which is large enough for @nr_groups groups containing
2433 * @nr_units units.  The returned ai's groups[0].cpu_map points to the
2434 * cpu_map array which is long enough for @nr_units and filled with
2435 * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
2436 * pointer of other groups.
2437 *
2438 * RETURNS:
2439 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2440 * failure.
2441 */
2442struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2443                                                      int nr_units)
2444{
2445        struct pcpu_alloc_info *ai;
2446        size_t base_size, ai_size;
2447        void *ptr;
2448        int unit;
2449
2450        base_size = ALIGN(struct_size(ai, groups, nr_groups),
2451                          __alignof__(ai->groups[0].cpu_map[0]));
2452        ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2453
2454        ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2455        if (!ptr)
2456                return NULL;
2457        ai = ptr;
2458        ptr += base_size;
2459
2460        ai->groups[0].cpu_map = ptr;
2461
2462        for (unit = 0; unit < nr_units; unit++)
2463                ai->groups[0].cpu_map[unit] = NR_CPUS;
2464
2465        ai->nr_groups = nr_groups;
2466        ai->__ai_size = PFN_ALIGN(ai_size);
2467
2468        return ai;
2469}
2470
2471/**
2472 * pcpu_free_alloc_info - free percpu allocation info
2473 * @ai: pcpu_alloc_info to free
2474 *
2475 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2476 */
2477void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2478{
2479        memblock_free_early(__pa(ai), ai->__ai_size);
2480}
2481
2482/**
2483 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2484 * @lvl: loglevel
2485 * @ai: allocation info to dump
2486 *
2487 * Print out information about @ai using loglevel @lvl.
2488 */
2489static void pcpu_dump_alloc_info(const char *lvl,
2490                                 const struct pcpu_alloc_info *ai)
2491{
2492        int group_width = 1, cpu_width = 1, width;
2493        char empty_str[] = "--------";
2494        int alloc = 0, alloc_end = 0;
2495        int group, v;
2496        int upa, apl;   /* units per alloc, allocs per line */
2497
2498        v = ai->nr_groups;
2499        while (v /= 10)
2500                group_width++;
2501
2502        v = num_possible_cpus();
2503        while (v /= 10)
2504                cpu_width++;
2505        empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2506
2507        upa = ai->alloc_size / ai->unit_size;
2508        width = upa * (cpu_width + 1) + group_width + 3;
2509        apl = rounddown_pow_of_two(max(60 / width, 1));
2510
2511        printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2512               lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2513               ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2514
2515        for (group = 0; group < ai->nr_groups; group++) {
2516                const struct pcpu_group_info *gi = &ai->groups[group];
2517                int unit = 0, unit_end = 0;
2518
2519                BUG_ON(gi->nr_units % upa);
2520                for (alloc_end += gi->nr_units / upa;
2521                     alloc < alloc_end; alloc++) {
2522                        if (!(alloc % apl)) {
2523                                pr_cont("\n");
2524                                printk("%spcpu-alloc: ", lvl);
2525                        }
2526                        pr_cont("[%0*d] ", group_width, group);
2527
2528                        for (unit_end += upa; unit < unit_end; unit++)
2529                                if (gi->cpu_map[unit] != NR_CPUS)
2530                                        pr_cont("%0*d ",
2531                                                cpu_width, gi->cpu_map[unit]);
2532                                else
2533                                        pr_cont("%s ", empty_str);
2534                }
2535        }
2536        pr_cont("\n");
2537}
2538
2539/**
2540 * pcpu_setup_first_chunk - initialize the first percpu chunk
2541 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2542 * @base_addr: mapped address
2543 *
2544 * Initialize the first percpu chunk which contains the kernel static
2545 * percpu area.  This function is to be called from arch percpu area
2546 * setup path.
2547 *
2548 * @ai contains all information necessary to initialize the first
2549 * chunk and prime the dynamic percpu allocator.
2550 *
2551 * @ai->static_size is the size of static percpu area.
2552 *
2553 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2554 * reserve after the static area in the first chunk.  This reserves
2555 * the first chunk such that it's available only through reserved
2556 * percpu allocation.  This is primarily used to serve module percpu
2557 * static areas on architectures where the addressing model has
2558 * limited offset range for symbol relocations to guarantee module
2559 * percpu symbols fall inside the relocatable range.
2560 *
2561 * @ai->dyn_size determines the number of bytes available for dynamic
2562 * allocation in the first chunk.  The area between @ai->static_size +
2563 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2564 *
2565 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2566 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2567 * @ai->dyn_size.
2568 *
2569 * @ai->atom_size is the allocation atom size and used as alignment
2570 * for vm areas.
2571 *
2572 * @ai->alloc_size is the allocation size and always multiple of
2573 * @ai->atom_size.  This is larger than @ai->atom_size if
2574 * @ai->unit_size is larger than @ai->atom_size.
2575 *
2576 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2577 * percpu areas.  Units which should be colocated are put into the
2578 * same group.  Dynamic VM areas will be allocated according to these
2579 * groupings.  If @ai->nr_groups is zero, a single group containing
2580 * all units is assumed.
2581 *
2582 * The caller should have mapped the first chunk at @base_addr and
2583 * copied static data to each unit.
2584 *
2585 * The first chunk will always contain a static and a dynamic region.
2586 * However, the static region is not managed by any chunk.  If the first
2587 * chunk also contains a reserved region, it is served by two chunks -
2588 * one for the reserved region and one for the dynamic region.  They
2589 * share the same vm, but use offset regions in the area allocation map.
2590 * The chunk serving the dynamic region is circulated in the chunk slots
2591 * and available for dynamic allocation like any other chunk.
2592 */
2593void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2594                                   void *base_addr)
2595{
2596        size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2597        size_t static_size, dyn_size;
2598        struct pcpu_chunk *chunk;
2599        unsigned long *group_offsets;
2600        size_t *group_sizes;
2601        unsigned long *unit_off;
2602        unsigned int cpu;
2603        int *unit_map;
2604        int group, unit, i;
2605        int map_size;
2606        unsigned long tmp_addr;
2607        size_t alloc_size;
2608
2609#define PCPU_SETUP_BUG_ON(cond) do {                                    \
2610        if (unlikely(cond)) {                                           \
2611                pr_emerg("failed to initialize, %s\n", #cond);          \
2612                pr_emerg("cpu_possible_mask=%*pb\n",                    \
2613                         cpumask_pr_args(cpu_possible_mask));           \
2614                pcpu_dump_alloc_info(KERN_EMERG, ai);                   \
2615                BUG();                                                  \
2616        }                                                               \
2617} while (0)
2618
2619        /* sanity checks */
2620        PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2621#ifdef CONFIG_SMP
2622        PCPU_SETUP_BUG_ON(!ai->static_size);
2623        PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2624#endif
2625        PCPU_SETUP_BUG_ON(!base_addr);
2626        PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2627        PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2628        PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2629        PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2630        PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2631        PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2632        PCPU_SETUP_BUG_ON(!ai->dyn_size);
2633        PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2634        PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2635                            IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2636        PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2637
2638        /* process group information and build config tables accordingly */
2639        alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2640        group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2641        if (!group_offsets)
2642                panic("%s: Failed to allocate %zu bytes\n", __func__,
2643                      alloc_size);
2644
2645        alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2646        group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2647        if (!group_sizes)
2648                panic("%s: Failed to allocate %zu bytes\n", __func__,
2649                      alloc_size);
2650
2651        alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2652        unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2653        if (!unit_map)
2654                panic("%s: Failed to allocate %zu bytes\n", __func__,
2655                      alloc_size);
2656
2657        alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2658        unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2659        if (!unit_off)
2660                panic("%s: Failed to allocate %zu bytes\n", __func__,
2661                      alloc_size);
2662
2663        for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2664                unit_map[cpu] = UINT_MAX;
2665
2666        pcpu_low_unit_cpu = NR_CPUS;
2667        pcpu_high_unit_cpu = NR_CPUS;
2668
2669        for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2670                const struct pcpu_group_info *gi = &ai->groups[group];
2671
2672                group_offsets[group] = gi->base_offset;
2673                group_sizes[group] = gi->nr_units * ai->unit_size;
2674
2675                for (i = 0; i < gi->nr_units; i++) {
2676                        cpu = gi->cpu_map[i];
2677                        if (cpu == NR_CPUS)
2678                                continue;
2679
2680                        PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2681                        PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2682                        PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2683
2684                        unit_map[cpu] = unit + i;
2685                        unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2686
2687                        /* determine low/high unit_cpu */
2688                        if (pcpu_low_unit_cpu == NR_CPUS ||
2689                            unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2690                                pcpu_low_unit_cpu = cpu;
2691                        if (pcpu_high_unit_cpu == NR_CPUS ||
2692                            unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2693                                pcpu_high_unit_cpu = cpu;
2694                }
2695        }
2696        pcpu_nr_units = unit;
2697
2698        for_each_possible_cpu(cpu)
2699                PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2700
2701        /* we're done parsing the input, undefine BUG macro and dump config */
2702#undef PCPU_SETUP_BUG_ON
2703        pcpu_dump_alloc_info(KERN_DEBUG, ai);
2704
2705        pcpu_nr_groups = ai->nr_groups;
2706        pcpu_group_offsets = group_offsets;
2707        pcpu_group_sizes = group_sizes;
2708        pcpu_unit_map = unit_map;
2709        pcpu_unit_offsets = unit_off;
2710
2711        /* determine basic parameters */
2712        pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2713        pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2714        pcpu_atom_size = ai->atom_size;
2715        pcpu_chunk_struct_size = struct_size(chunk, populated,
2716                                             BITS_TO_LONGS(pcpu_unit_pages));
2717
2718        pcpu_stats_save_ai(ai);
2719
2720        /*
2721         * Allocate chunk slots.  The slots after the active slots are:
2722         *   sidelined_slot - isolated, depopulated chunks
2723         *   free_slot - fully free chunks
2724         *   to_depopulate_slot - isolated, chunks to depopulate
2725         */
2726        pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1;
2727        pcpu_free_slot = pcpu_sidelined_slot + 1;
2728        pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2729        pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2730        pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2731                                          sizeof(pcpu_chunk_lists[0]),
2732                                          SMP_CACHE_BYTES);
2733        if (!pcpu_chunk_lists)
2734                panic("%s: Failed to allocate %zu bytes\n", __func__,
2735                      pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2736
2737        for (i = 0; i < pcpu_nr_slots; i++)
2738                INIT_LIST_HEAD(&pcpu_chunk_lists[i]);
2739
2740        /*
2741         * The end of the static region needs to be aligned with the
2742         * minimum allocation size as this offsets the reserved and
2743         * dynamic region.  The first chunk ends page aligned by
2744         * expanding the dynamic region, therefore the dynamic region
2745         * can be shrunk to compensate while still staying above the
2746         * configured sizes.
2747         */
2748        static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2749        dyn_size = ai->dyn_size - (static_size - ai->static_size);
2750
2751        /*
2752         * Initialize first chunk.
2753         * If the reserved_size is non-zero, this initializes the reserved
2754         * chunk.  If the reserved_size is zero, the reserved chunk is NULL
2755         * and the dynamic region is initialized here.  The first chunk,
2756         * pcpu_first_chunk, will always point to the chunk that serves
2757         * the dynamic region.
2758         */
2759        tmp_addr = (unsigned long)base_addr + static_size;
2760        map_size = ai->reserved_size ?: dyn_size;
2761        chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2762
2763        /* init dynamic chunk if necessary */
2764        if (ai->reserved_size) {
2765                pcpu_reserved_chunk = chunk;
2766
2767                tmp_addr = (unsigned long)base_addr + static_size +
2768                           ai->reserved_size;
2769                map_size = dyn_size;
2770                chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2771        }
2772
2773        /* link the first chunk in */
2774        pcpu_first_chunk = chunk;
2775        pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2776        pcpu_chunk_relocate(pcpu_first_chunk, -1);
2777
2778        /* include all regions of the first chunk */
2779        pcpu_nr_populated += PFN_DOWN(size_sum);
2780
2781        pcpu_stats_chunk_alloc();
2782        trace_percpu_create_chunk(base_addr);
2783
2784        /* we're done */
2785        pcpu_base_addr = base_addr;
2786}
2787
2788#ifdef CONFIG_SMP
2789
2790const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2791        [PCPU_FC_AUTO]  = "auto",
2792        [PCPU_FC_EMBED] = "embed",
2793        [PCPU_FC_PAGE]  = "page",
2794};
2795
2796enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2797
2798static int __init percpu_alloc_setup(char *str)
2799{
2800        if (!str)
2801                return -EINVAL;
2802
2803        if (0)
2804                /* nada */;
2805#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2806        else if (!strcmp(str, "embed"))
2807                pcpu_chosen_fc = PCPU_FC_EMBED;
2808#endif
2809#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2810        else if (!strcmp(str, "page"))
2811                pcpu_chosen_fc = PCPU_FC_PAGE;
2812#endif
2813        else
2814                pr_warn("unknown allocator %s specified\n", str);
2815
2816        return 0;
2817}
2818early_param("percpu_alloc", percpu_alloc_setup);
2819
2820/*
2821 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2822 * Build it if needed by the arch config or the generic setup is going
2823 * to be used.
2824 */
2825#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2826        !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2827#define BUILD_EMBED_FIRST_CHUNK
2828#endif
2829
2830/* build pcpu_page_first_chunk() iff needed by the arch config */
2831#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2832#define BUILD_PAGE_FIRST_CHUNK
2833#endif
2834
2835/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2836#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2837/**
2838 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2839 * @reserved_size: the size of reserved percpu area in bytes
2840 * @dyn_size: minimum free size for dynamic allocation in bytes
2841 * @atom_size: allocation atom size
2842 * @cpu_distance_fn: callback to determine distance between cpus, optional
2843 *
2844 * This function determines grouping of units, their mappings to cpus
2845 * and other parameters considering needed percpu size, allocation
2846 * atom size and distances between CPUs.
2847 *
2848 * Groups are always multiples of atom size and CPUs which are of
2849 * LOCAL_DISTANCE both ways are grouped together and share space for
2850 * units in the same group.  The returned configuration is guaranteed
2851 * to have CPUs on different nodes on different groups and >=75% usage
2852 * of allocated virtual address space.
2853 *
2854 * RETURNS:
2855 * On success, pointer to the new allocation_info is returned.  On
2856 * failure, ERR_PTR value is returned.
2857 */
2858static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2859                                size_t reserved_size, size_t dyn_size,
2860                                size_t atom_size,
2861                                pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2862{
2863        static int group_map[NR_CPUS] __initdata;
2864        static int group_cnt[NR_CPUS] __initdata;
2865        static struct cpumask mask __initdata;
2866        const size_t static_size = __per_cpu_end - __per_cpu_start;
2867        int nr_groups = 1, nr_units = 0;
2868        size_t size_sum, min_unit_size, alloc_size;
2869        int upa, max_upa, best_upa;     /* units_per_alloc */
2870        int last_allocs, group, unit;
2871        unsigned int cpu, tcpu;
2872        struct pcpu_alloc_info *ai;
2873        unsigned int *cpu_map;
2874
2875        /* this function may be called multiple times */
2876        memset(group_map, 0, sizeof(group_map));
2877        memset(group_cnt, 0, sizeof(group_cnt));
2878        cpumask_clear(&mask);
2879
2880        /* calculate size_sum and ensure dyn_size is enough for early alloc */
2881        size_sum = PFN_ALIGN(static_size + reserved_size +
2882                            max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2883        dyn_size = size_sum - static_size - reserved_size;
2884
2885        /*
2886         * Determine min_unit_size, alloc_size and max_upa such that
2887         * alloc_size is multiple of atom_size and is the smallest
2888         * which can accommodate 4k aligned segments which are equal to
2889         * or larger than min_unit_size.
2890         */
2891        min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2892
2893        /* determine the maximum # of units that can fit in an allocation */
2894        alloc_size = roundup(min_unit_size, atom_size);
2895        upa = alloc_size / min_unit_size;
2896        while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2897                upa--;
2898        max_upa = upa;
2899
2900        cpumask_copy(&mask, cpu_possible_mask);
2901
2902        /* group cpus according to their proximity */
2903        for (group = 0; !cpumask_empty(&mask); group++) {
2904                /* pop the group's first cpu */
2905                cpu = cpumask_first(&mask);
2906                group_map[cpu] = group;
2907                group_cnt[group]++;
2908                cpumask_clear_cpu(cpu, &mask);
2909
2910                for_each_cpu(tcpu, &mask) {
2911                        if (!cpu_distance_fn ||
2912                            (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2913                             cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2914                                group_map[tcpu] = group;
2915                                group_cnt[group]++;
2916                                cpumask_clear_cpu(tcpu, &mask);
2917                        }
2918                }
2919        }
2920        nr_groups = group;
2921
2922        /*
2923         * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2924         * Expand the unit_size until we use >= 75% of the units allocated.
2925         * Related to atom_size, which could be much larger than the unit_size.
2926         */
2927        last_allocs = INT_MAX;
2928        best_upa = 0;
2929        for (upa = max_upa; upa; upa--) {
2930                int allocs = 0, wasted = 0;
2931
2932                if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2933                        continue;
2934
2935                for (group = 0; group < nr_groups; group++) {
2936                        int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2937                        allocs += this_allocs;
2938                        wasted += this_allocs * upa - group_cnt[group];
2939                }
2940
2941                /*
2942                 * Don't accept if wastage is over 1/3.  The
2943                 * greater-than comparison ensures upa==1 always
2944                 * passes the following check.
2945                 */
2946                if (wasted > num_possible_cpus() / 3)
2947                        continue;
2948
2949                /* and then don't consume more memory */
2950                if (allocs > last_allocs)
2951                        break;
2952                last_allocs = allocs;
2953                best_upa = upa;
2954        }
2955        BUG_ON(!best_upa);
2956        upa = best_upa;
2957
2958        /* allocate and fill alloc_info */
2959        for (group = 0; group < nr_groups; group++)
2960                nr_units += roundup(group_cnt[group], upa);
2961
2962        ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2963        if (!ai)
2964                return ERR_PTR(-ENOMEM);
2965        cpu_map = ai->groups[0].cpu_map;
2966
2967        for (group = 0; group < nr_groups; group++) {
2968                ai->groups[group].cpu_map = cpu_map;
2969                cpu_map += roundup(group_cnt[group], upa);
2970        }
2971
2972        ai->static_size = static_size;
2973        ai->reserved_size = reserved_size;
2974        ai->dyn_size = dyn_size;
2975        ai->unit_size = alloc_size / upa;
2976        ai->atom_size = atom_size;
2977        ai->alloc_size = alloc_size;
2978
2979        for (group = 0, unit = 0; group < nr_groups; group++) {
2980                struct pcpu_group_info *gi = &ai->groups[group];
2981
2982                /*
2983                 * Initialize base_offset as if all groups are located
2984                 * back-to-back.  The caller should update this to
2985                 * reflect actual allocation.
2986                 */
2987                gi->base_offset = unit * ai->unit_size;
2988
2989                for_each_possible_cpu(cpu)
2990                        if (group_map[cpu] == group)
2991                                gi->cpu_map[gi->nr_units++] = cpu;
2992                gi->nr_units = roundup(gi->nr_units, upa);
2993                unit += gi->nr_units;
2994        }
2995        BUG_ON(unit != nr_units);
2996
2997        return ai;
2998}
2999#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
3000
3001#if defined(BUILD_EMBED_FIRST_CHUNK)
3002/**
3003 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
3004 * @reserved_size: the size of reserved percpu area in bytes
3005 * @dyn_size: minimum free size for dynamic allocation in bytes
3006 * @atom_size: allocation atom size
3007 * @cpu_distance_fn: callback to determine distance between cpus, optional
3008 * @alloc_fn: function to allocate percpu page
3009 * @free_fn: function to free percpu page
3010 *
3011 * This is a helper to ease setting up embedded first percpu chunk and
3012 * can be called where pcpu_setup_first_chunk() is expected.
3013 *
3014 * If this function is used to setup the first chunk, it is allocated
3015 * by calling @alloc_fn and used as-is without being mapped into
3016 * vmalloc area.  Allocations are always whole multiples of @atom_size
3017 * aligned to @atom_size.
3018 *
3019 * This enables the first chunk to piggy back on the linear physical
3020 * mapping which often uses larger page size.  Please note that this
3021 * can result in very sparse cpu->unit mapping on NUMA machines thus
3022 * requiring large vmalloc address space.  Don't use this allocator if
3023 * vmalloc space is not orders of magnitude larger than distances
3024 * between node memory addresses (ie. 32bit NUMA machines).
3025 *
3026 * @dyn_size specifies the minimum dynamic area size.
3027 *
3028 * If the needed size is smaller than the minimum or specified unit
3029 * size, the leftover is returned using @free_fn.
3030 *
3031 * RETURNS:
3032 * 0 on success, -errno on failure.
3033 */
3034int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3035                                  size_t atom_size,
3036                                  pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3037                                  pcpu_fc_alloc_fn_t alloc_fn,
3038                                  pcpu_fc_free_fn_t free_fn)
3039{
3040        void *base = (void *)ULONG_MAX;
3041        void **areas = NULL;
3042        struct pcpu_alloc_info *ai;
3043        size_t size_sum, areas_size;
3044        unsigned long max_distance;
3045        int group, i, highest_group, rc = 0;
3046
3047        ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3048                                   cpu_distance_fn);
3049        if (IS_ERR(ai))
3050                return PTR_ERR(ai);
3051
3052        size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3053        areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3054
3055        areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
3056        if (!areas) {
3057                rc = -ENOMEM;
3058                goto out_free;
3059        }
3060
3061        /* allocate, copy and determine base address & max_distance */
3062        highest_group = 0;
3063        for (group = 0; group < ai->nr_groups; group++) {
3064                struct pcpu_group_info *gi = &ai->groups[group];
3065                unsigned int cpu = NR_CPUS;
3066                void *ptr;
3067
3068                for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3069                        cpu = gi->cpu_map[i];
3070                BUG_ON(cpu == NR_CPUS);
3071
3072                /* allocate space for the whole group */
3073                ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
3074                if (!ptr) {
3075                        rc = -ENOMEM;
3076                        goto out_free_areas;
3077                }
3078                /* kmemleak tracks the percpu allocations separately */
3079                kmemleak_free(ptr);
3080                areas[group] = ptr;
3081
3082                base = min(ptr, base);
3083                if (ptr > areas[highest_group])
3084                        highest_group = group;
3085        }
3086        max_distance = areas[highest_group] - base;
3087        max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3088
3089        /* warn if maximum distance is further than 75% of vmalloc space */
3090        if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3091                pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3092                                max_distance, VMALLOC_TOTAL);
3093#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3094                /* and fail if we have fallback */
3095                rc = -EINVAL;
3096                goto out_free_areas;
3097#endif
3098        }
3099
3100        /*
3101         * Copy data and free unused parts.  This should happen after all
3102         * allocations are complete; otherwise, we may end up with
3103         * overlapping groups.
3104         */
3105        for (group = 0; group < ai->nr_groups; group++) {
3106                struct pcpu_group_info *gi = &ai->groups[group];
3107                void *ptr = areas[group];
3108
3109                for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3110                        if (gi->cpu_map[i] == NR_CPUS) {
3111                                /* unused unit, free whole */
3112                                free_fn(ptr, ai->unit_size);
3113                                continue;
3114                        }
3115                        /* copy and return the unused part */
3116                        memcpy(ptr, __per_cpu_load, ai->static_size);
3117                        free_fn(ptr + size_sum, ai->unit_size - size_sum);
3118                }
3119        }
3120
3121        /* base address is now known, determine group base offsets */
3122        for (group = 0; group < ai->nr_groups; group++) {
3123                ai->groups[group].base_offset = areas[group] - base;
3124        }
3125
3126        pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3127                PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3128                ai->dyn_size, ai->unit_size);
3129
3130        pcpu_setup_first_chunk(ai, base);
3131        goto out_free;
3132
3133out_free_areas:
3134        for (group = 0; group < ai->nr_groups; group++)
3135                if (areas[group])
3136                        free_fn(areas[group],
3137                                ai->groups[group].nr_units * ai->unit_size);
3138out_free:
3139        pcpu_free_alloc_info(ai);
3140        if (areas)
3141                memblock_free_early(__pa(areas), areas_size);
3142        return rc;
3143}
3144#endif /* BUILD_EMBED_FIRST_CHUNK */
3145
3146#ifdef BUILD_PAGE_FIRST_CHUNK
3147/**
3148 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3149 * @reserved_size: the size of reserved percpu area in bytes
3150 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
3151 * @free_fn: function to free percpu page, always called with PAGE_SIZE
3152 * @populate_pte_fn: function to populate pte
3153 *
3154 * This is a helper to ease setting up page-remapped first percpu
3155 * chunk and can be called where pcpu_setup_first_chunk() is expected.
3156 *
3157 * This is the basic allocator.  Static percpu area is allocated
3158 * page-by-page into vmalloc area.
3159 *
3160 * RETURNS:
3161 * 0 on success, -errno on failure.
3162 */
3163int __init pcpu_page_first_chunk(size_t reserved_size,
3164                                 pcpu_fc_alloc_fn_t alloc_fn,
3165                                 pcpu_fc_free_fn_t free_fn,
3166                                 pcpu_fc_populate_pte_fn_t populate_pte_fn)
3167{
3168        static struct vm_struct vm;
3169        struct pcpu_alloc_info *ai;
3170        char psize_str[16];
3171        int unit_pages;
3172        size_t pages_size;
3173        struct page **pages;
3174        int unit, i, j, rc = 0;
3175        int upa;
3176        int nr_g0_units;
3177
3178        snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
3179
3180        ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
3181        if (IS_ERR(ai))
3182                return PTR_ERR(ai);
3183        BUG_ON(ai->nr_groups != 1);
3184        upa = ai->alloc_size/ai->unit_size;
3185        nr_g0_units = roundup(num_possible_cpus(), upa);
3186        if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3187                pcpu_free_alloc_info(ai);
3188                return -EINVAL;
3189        }
3190
3191        unit_pages = ai->unit_size >> PAGE_SHIFT;
3192
3193        /* unaligned allocations can't be freed, round up to page size */
3194        pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3195                               sizeof(pages[0]));
3196        pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
3197        if (!pages)
3198                panic("%s: Failed to allocate %zu bytes\n", __func__,
3199                      pages_size);
3200
3201        /* allocate pages */
3202        j = 0;
3203        for (unit = 0; unit < num_possible_cpus(); unit++) {
3204                unsigned int cpu = ai->groups[0].cpu_map[unit];
3205                for (i = 0; i < unit_pages; i++) {
3206                        void *ptr;
3207
3208                        ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
3209                        if (!ptr) {
3210                                pr_warn("failed to allocate %s page for cpu%u\n",
3211                                                psize_str, cpu);
3212                                goto enomem;
3213                        }
3214                        /* kmemleak tracks the percpu allocations separately */
3215                        kmemleak_free(ptr);
3216                        pages[j++] = virt_to_page(ptr);
3217                }
3218        }
3219
3220        /* allocate vm area, map the pages and copy static data */
3221        vm.flags = VM_ALLOC;
3222        vm.size = num_possible_cpus() * ai->unit_size;
3223        vm_area_register_early(&vm, PAGE_SIZE);
3224
3225        for (unit = 0; unit < num_possible_cpus(); unit++) {
3226                unsigned long unit_addr =
3227                        (unsigned long)vm.addr + unit * ai->unit_size;
3228
3229                for (i = 0; i < unit_pages; i++)
3230                        populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
3231
3232                /* pte already populated, the following shouldn't fail */
3233                rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3234                                      unit_pages);
3235                if (rc < 0)
3236                        panic("failed to map percpu area, err=%d\n", rc);
3237
3238                /*
3239                 * FIXME: Archs with virtual cache should flush local
3240                 * cache for the linear mapping here - something
3241                 * equivalent to flush_cache_vmap() on the local cpu.
3242                 * flush_cache_vmap() can't be used as most supporting
3243                 * data structures are not set up yet.
3244                 */
3245
3246                /* copy static data */
3247                memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3248        }
3249
3250        /* we're ready, commit */
3251        pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3252                unit_pages, psize_str, ai->static_size,
3253                ai->reserved_size, ai->dyn_size);
3254
3255        pcpu_setup_first_chunk(ai, vm.addr);
3256        goto out_free_ar;
3257
3258enomem:
3259        while (--j >= 0)
3260                free_fn(page_address(pages[j]), PAGE_SIZE);
3261        rc = -ENOMEM;
3262out_free_ar:
3263        memblock_free_early(__pa(pages), pages_size);
3264        pcpu_free_alloc_info(ai);
3265        return rc;
3266}
3267#endif /* BUILD_PAGE_FIRST_CHUNK */
3268
3269#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3270/*
3271 * Generic SMP percpu area setup.
3272 *
3273 * The embedding helper is used because its behavior closely resembles
3274 * the original non-dynamic generic percpu area setup.  This is
3275 * important because many archs have addressing restrictions and might
3276 * fail if the percpu area is located far away from the previous
3277 * location.  As an added bonus, in non-NUMA cases, embedding is
3278 * generally a good idea TLB-wise because percpu area can piggy back
3279 * on the physical linear memory mapping which uses large page
3280 * mappings on applicable archs.
3281 */
3282unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3283EXPORT_SYMBOL(__per_cpu_offset);
3284
3285static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
3286                                       size_t align)
3287{
3288        return  memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS));
3289}
3290
3291static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
3292{
3293        memblock_free_early(__pa(ptr), size);
3294}
3295
3296void __init setup_per_cpu_areas(void)
3297{
3298        unsigned long delta;
3299        unsigned int cpu;
3300        int rc;
3301
3302        /*
3303         * Always reserve area for module percpu variables.  That's
3304         * what the legacy allocator did.
3305         */
3306        rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
3307                                    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
3308                                    pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
3309        if (rc < 0)
3310                panic("Failed to initialize percpu areas.");
3311
3312        delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3313        for_each_possible_cpu(cpu)
3314                __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3315}
3316#endif  /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3317
3318#else   /* CONFIG_SMP */
3319
3320/*
3321 * UP percpu area setup.
3322 *
3323 * UP always uses km-based percpu allocator with identity mapping.
3324 * Static percpu variables are indistinguishable from the usual static
3325 * variables and don't require any special preparation.
3326 */
3327void __init setup_per_cpu_areas(void)
3328{
3329        const size_t unit_size =
3330                roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3331                                         PERCPU_DYNAMIC_RESERVE));
3332        struct pcpu_alloc_info *ai;
3333        void *fc;
3334
3335        ai = pcpu_alloc_alloc_info(1, 1);
3336        fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3337        if (!ai || !fc)
3338                panic("Failed to allocate memory for percpu areas.");
3339        /* kmemleak tracks the percpu allocations separately */
3340        kmemleak_free(fc);
3341
3342        ai->dyn_size = unit_size;
3343        ai->unit_size = unit_size;
3344        ai->atom_size = unit_size;
3345        ai->alloc_size = unit_size;
3346        ai->groups[0].nr_units = 1;
3347        ai->groups[0].cpu_map[0] = 0;
3348
3349        pcpu_setup_first_chunk(ai, fc);
3350        pcpu_free_alloc_info(ai);
3351}
3352
3353#endif  /* CONFIG_SMP */
3354
3355/*
3356 * pcpu_nr_pages - calculate total number of populated backing pages
3357 *
3358 * This reflects the number of pages populated to back chunks.  Metadata is
3359 * excluded in the number exposed in meminfo as the number of backing pages
3360 * scales with the number of cpus and can quickly outweigh the memory used for
3361 * metadata.  It also keeps this calculation nice and simple.
3362 *
3363 * RETURNS:
3364 * Total number of populated backing pages in use by the allocator.
3365 */
3366unsigned long pcpu_nr_pages(void)
3367{
3368        return pcpu_nr_populated * pcpu_nr_units;
3369}
3370
3371/*
3372 * Percpu allocator is initialized early during boot when neither slab or
3373 * workqueue is available.  Plug async management until everything is up
3374 * and running.
3375 */
3376static int __init percpu_enable_async(void)
3377{
3378        pcpu_async_enabled = true;
3379        return 0;
3380}
3381subsys_initcall(percpu_enable_async);
3382
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.