Objective-C RunTime内存

iOS 高级之美(六)—— malloc分析

2019-12-29  本文已影响0人  Cooci_和谐学习_不急不躁

前言: iOS 高级之美 是本人总结了一些工作实际开发研究以及面试重点,围绕底层进行 源码分析 - LLDB 调试 - 源码断点 - 汇编调试,让读者真正感受 Runtime底层之美~😊
目录如下:

iOS 高级之美(一)—— iOS_objc4-756.2 最新源码编译调试
iOS 高级之美(二)—— OC对象底层上篇
iOS 高级之美(三)—— OC对象底层下篇
iOS 高级之美(四)—— isa原理分析
iOS 高级之美(五)—— 类结构分析
iOS 高级之美(六)—— malloc分析

image

我们前面分析了对象的创建,其中一个非常重要的点:申请内存空间!
然而 obj = (id)calloc(1, size) 这一段代码所在位置不再是 libObjc4,它定位到了 libmalloc , 至于愈合定位的大家可以参考笔者前面的文章。这个篇章我们针对 malloc 展开而分析

那么 calloc 方法做了什么呢,让我们来一探究竟!

一、malloc_zone_t 分析

这个家伙是一个非常重要的家伙,我们先来看看 malloc_zone_t 的结构

typedef struct _malloc_zone_t {
void    *reserved1;    /* RESERVED FOR CFAllocator DO NOT USE */
    void    *reserved2;    /* RESERVED FOR CFAllocator DO NOT USE */
    size_t     (* MALLOC_ZONE_FN_PTR(size))(struct _malloc_zone_t *zone, const void *ptr); /* returns the size of a block or 0 if not in this zone; must be fast, especially for negative answers */
    void     *(* MALLOC_ZONE_FN_PTR(malloc))(struct _malloc_zone_t *zone, size_t size);
    void     *(* MALLOC_ZONE_FN_PTR(calloc))(struct _malloc_zone_t *zone, size_t num_items, size_t size); /* same as malloc, but block returned is set to zero */
    void     *(* MALLOC_ZONE_FN_PTR(valloc))(struct _malloc_zone_t *zone, size_t size); /* same as malloc, but block returned is set to zero and is guaranteed to be page aligned */
    void     (* MALLOC_ZONE_FN_PTR(free))(struct _malloc_zone_t *zone, void *ptr);
    void     *(* MALLOC_ZONE_FN_PTR(realloc))(struct _malloc_zone_t *zone, void *ptr, size_t size);
    void     (* MALLOC_ZONE_FN_PTR(destroy))(struct _malloc_zone_t *zone);
    const char    *zone_name;

    unsigned    (* MALLOC_ZONE_FN_PTR(batch_malloc))(struct _malloc_zone_t *zone, size_t size, void **results, unsigned num_requested);

    struct malloc_introspection_t    * MALLOC_INTROSPECT_TBL_PTR(introspect);
    unsigned    version;

    void *(* MALLOC_ZONE_FN_PTR(memalign))(struct _malloc_zone_t *zone, size_t alignment, size_t size);

    void (* MALLOC_ZONE_FN_PTR(free_definite_size))(struct _malloc_zone_t *zone, void *ptr, size_t size);

    size_t     (* MALLOC_ZONE_FN_PTR(pressure_relief))(struct _malloc_zone_t *zone, size_t goal);

    boolean_t (* MALLOC_ZONE_FN_PTR(claimed_address))(struct _malloc_zone_t *zone, void *ptr);

} malloc_zone_t;

malloc_zone_t 是一个非常基础结构,里面包含一堆函数指针,用来存储一堆相关的处理函数的具体实现的地址,例如mallocfreerealloc等函数的具体实现。后续会基于malloc_zone_t进行扩展。

二、calloc 的流程

2.1 calloc -> malloc_zone_calloc 的流程

void * calloc(size_t num_items, size_t size)
{
    void *retval;
    retval = malloc_zone_calloc(default_zone, num_items, size);
    if (retval == NULL) {
        errno = ENOMEM;
    }
    return retval;
}

2.2 default_zone 引导

void * malloc_zone_calloc(malloc_zone_t *zone, size_t num_items, size_t size)
{
    MALLOC_TRACE(TRACE_calloc | DBG_FUNC_START, (uintptr_t)zone, num_items, size, 0);

    void *ptr;
    if (malloc_check_start && (malloc_check_counter++ >= malloc_check_start)) {
        internal_check();
    }

    ptr = zone->calloc(zone, num_items, size);
    
    if (malloc_logger) {
        malloc_logger(MALLOC_LOG_TYPE_ALLOCATE | MALLOC_LOG_TYPE_HAS_ZONE | MALLOC_LOG_TYPE_CLEARED, (uintptr_t)zone,
                (uintptr_t)(num_items * size), 0, (uintptr_t)ptr, 0);
    }

    MALLOC_TRACE(TRACE_calloc | DBG_FUNC_END, (uintptr_t)zone, num_items, size, (uintptr_t)ptr);
    return ptr;
}

2.3 defaultzone 的定义

static virtual_default_zone_t virtual_default_zone
__attribute__((section("__DATA,__v_zone")))
__attribute__((aligned(PAGE_MAX_SIZE))) = {
    NULL,
    NULL,
    default_zone_size,
    default_zone_malloc,
    default_zone_calloc,
    default_zone_valloc,
    default_zone_free,
    default_zone_realloc,
    default_zone_destroy,
    DEFAULT_MALLOC_ZONE_STRING,
    default_zone_batch_malloc,
    default_zone_batch_free,
    &default_zone_introspect,
    10,
    default_zone_memalign,
    default_zone_free_definite_size,
    default_zone_pressure_relief,
    default_zone_malloc_claimed_address,
};
static void *
default_zone_calloc(malloc_zone_t *zone, size_t num_items, size_t size)
{
    zone = runtime_default_zone();
    
    return zone->calloc(zone, num_items, size);
}

2.4 zone分析

在创建正在的 zone时,其实系统是有对应的一套创建策略的。在跟踪 runtime_default_zone 方法后,最终会进入如下调用

image
static void
_malloc_initialize(void *context __unused)
{
    ...... - 省略多余代码
    //创建helper_zone,
    malloc_zone_t *helper_zone = create_scalable_zone(0, malloc_debug_flags);
    //创建 nano zone
    if (_malloc_engaged_nano == NANO_V2) {
    zone = nanov2_create_zone(helper_zone, malloc_debug_flags);
    } else if (_malloc_engaged_nano == NANO_V1) {
    zone = nano_create_zone(helper_zone, malloc_debug_flags);
    }
    //如果上面的if else if 成立,这进入 nonazone
    if (zone) {
    malloc_zone_register_while_locked(zone);
    malloc_zone_register_while_locked(helper_zone);

    // Must call malloc_set_zone_name() *after* helper and nano are hooked together.
    malloc_set_zone_name(zone, DEFAULT_MALLOC_ZONE_STRING);
    malloc_set_zone_name(helper_zone, MALLOC_HELPER_ZONE_STRING);
    } else {
    //使用helper_zone分配内存
    zone = helper_zone;
    malloc_zone_register_while_locked(zone);
    malloc_set_zone_name(zone, DEFAULT_MALLOC_ZONE_STRING);
    }
    //缓存default_zone
    initial_default_zone = zone;
    .....    
}

在这里 会存在两种 zone

2.5 nanozone_t 分析

typedef struct nano_meta_s {
 OSQueueHead            slot_LIFO MALLOC_NANO_CACHE_ALIGN;
    unsigned int        slot_madvised_log_page_count;
    volatile uintptr_t        slot_current_base_addr;
    volatile uintptr_t        slot_limit_addr;
    volatile size_t        slot_objects_mapped;
    volatile size_t        slot_objects_skipped;
    bitarray_t            slot_madvised_pages;
    // position on cache line distinct from that of slot_LIFO
    volatile uintptr_t        slot_bump_addr MALLOC_NANO_CACHE_ALIGN;
    volatile boolean_t        slot_exhausted;
    unsigned int        slot_bytes;
    unsigned int        slot_objects;
} *nano_meta_admin_t;

    // vm_allocate()'d, so page-aligned to begin with.
typedef struct nanozone_s {
    // first page will be given read-only protection
    malloc_zone_t        basic_zone;
    uint8_t            pad[PAGE_MAX_SIZE - sizeof(malloc_zone_t)];

    // remainder of structure is R/W (contains no function pointers)
    // page-aligned
    // max: NANO_MAG_SIZE cores x NANO_SLOT_SIZE slots for nano blocks {16 .. 256}
    //以Mag、Slot为维度,维护申请的band内存部分 slot 的范围为 1~16
    struct nano_meta_s        meta_data[NANO_MAG_SIZE][NANO_SLOT_SIZE];//
    _malloc_lock_s            band_resupply_lock[NANO_MAG_SIZE];
    uintptr_t           band_max_mapped_baseaddr[NANO_MAG_SIZE];
    size_t            core_mapped_size[NANO_MAG_SIZE];
    unsigned            debug_flags;
    uintptr_t            cookie;
    malloc_zone_t        *helper_zone;
} nanozone_t;

2.6 nano_create_zone 分析

malloc_zone_t *
nano_create_zone(malloc_zone_t *helper_zone, unsigned debug_flags)
{
    nanozone_t *nanozone;
    int i, j;
    //构造nano zone
    /* Note: It is important that nano_create_zone resets _malloc_engaged_nano
     * if it is unable to enable the nanozone (and chooses not to abort). As
     * several functions rely on _malloc_engaged_nano to determine if they
     * should manipulate the nanozone, and these should not run if we failed
     * to create the zone.
     */
//     MALLOC_ASSERT(_malloc_engaged_nano == NANO_V1);

    /* get memory for the zone. */
    nanozone = nano_common_allocate_based_pages(NANOZONE_PAGED_SIZE, 0, 0, VM_MEMORY_MALLOC, 0);
    if (!nanozone) {
        _malloc_engaged_nano = NANO_NONE;
        return NULL;
    }
    //构造对zone 的一些函数进行重新赋值
    /* set up the basic_zone portion of the nanozone structure */
    nanozone->basic_zone.version = 10;
    nanozone->basic_zone.size = (void *)nano_size;
    nanozone->basic_zone.malloc = (debug_flags & MALLOC_DO_SCRIBBLE) ? (void *)nano_malloc_scribble : (void *)nano_malloc;
    nanozone->basic_zone.calloc = (void *)nano_calloc;
    nanozone->basic_zone.valloc = (void *)nano_valloc;
    nanozone->basic_zone.free = (debug_flags & MALLOC_DO_SCRIBBLE) ? (void *)nano_free_scribble : (void *)nano_free;
    nanozone->basic_zone.realloc = (void *)nano_realloc;
    nanozone->basic_zone.destroy = (void *)nano_destroy;
    nanozone->basic_zone.batch_malloc = (void *)nano_batch_malloc;
    nanozone->basic_zone.batch_free = (void *)nano_batch_free;
    nanozone->basic_zone.introspect = (struct malloc_introspection_t *)&nano_introspect;
    nanozone->basic_zone.memalign = (void *)nano_memalign;
    nanozone->basic_zone.free_definite_size = (debug_flags & MALLOC_DO_SCRIBBLE) ? (void *)nano_free_definite_size_scribble
                                                                                          : (void *)nano_free_definite_size;

    nanozone->basic_zone.pressure_relief = (void *)nano_pressure_relief;
    nanozone->basic_zone.claimed_address = (void *)nano_claimed_address;

    nanozone->basic_zone.reserved1 = 0; /* Set to zero once and for all as required by CFAllocator. */
    nanozone->basic_zone.reserved2 = 0; /* Set to zero once and for all as required by CFAllocator. */

    mprotect(nanozone, sizeof(nanozone->basic_zone), PROT_READ); /* Prevent overwriting the function pointers in basic_zone. */

    /* Nano zone does not support MALLOC_ADD_GUARD_PAGES. */
    if (debug_flags & MALLOC_ADD_GUARD_PAGES) {
        malloc_report(ASL_LEVEL_INFO, "nano zone does not support guard pages\n");
        debug_flags &= ~MALLOC_ADD_GUARD_PAGES;
    }

    /* set up the remainder of the nanozone structure */
    nanozone->debug_flags = debug_flags;

    if (phys_ncpus > sizeof(nanozone->core_mapped_size) /
            sizeof(nanozone->core_mapped_size[0])) {
        MALLOC_REPORT_FATAL_ERROR(phys_ncpus,
                "nanozone abandoned because NCPUS > max magazines.\n");
    }

    /* Initialize slot queue heads and resupply locks. */
    OSQueueHead q0 = OS_ATOMIC_QUEUE_INIT;
    for (i = 0; i < nano_common_max_magazines; ++i) {
        _malloc_lock_init(&nanozone->band_resupply_lock[i]);

        for (j = 0; j < NANO_SLOT_SIZE; ++j) {
            nanozone->meta_data[i][j].slot_LIFO = q0;
        }
    }

    /* Initialize the security token. */
    nanozone->cookie = (uintptr_t)malloc_entropy[0] & 0x0000ffffffff0000ULL; // scramble central 32bits with this cookie

    nanozone->helper_zone = helper_zone;

    return (malloc_zone_t *)nanozone;
}

2.7 nano_calloc 分析

过程参考 defaultzone 。回到上面 default_zone_calloc 函数内。下一步就是使用 nanozone_t 调用 calloc

下面是 nano_calloc 的实现

static void *
nano_calloc(nanozone_t *nanozone, size_t num_items, size_t size)
{
    size_t total_bytes;

    if (calloc_get_size(num_items, size, 0, &total_bytes)) {
        return NULL;
    }
    // 如果要开辟的空间小于 NANO_MAX_SIZE 则进行nanozone_t的malloc。
    if (total_bytes <= NANO_MAX_SIZE) {
        void *p = _nano_malloc_check_clear(nanozone, total_bytes, 1);
        if (p) {
            return p;
        } else {
            /* FALLTHROUGH to helper zone */
        }
    }
    //否则就进行helper_zone的流程
    malloc_zone_t *zone = (malloc_zone_t *)(nanozone->helper_zone);
    return zone->calloc(zone, 1, total_bytes);
}

2.8 _nano_malloc_check_clear分析

这里我们也可以看出使用 nanozone_t 的限制为不超过256B。继续看 _nano_malloc_check_clear

static void *
_nano_malloc_check_clear(nanozone_t *nanozone, size_t size, boolean_t cleared_requested)
{
    MALLOC_TRACE(TRACE_nano_malloc, (uintptr_t)nanozone, size, cleared_requested, 0);

    void *ptr;
    size_t slot_key;
    // 获取16字节对齐之后的大小,slot_key非常关键,为slot_bytes/16的值,也是数组的二维下下标
    size_t slot_bytes = segregated_size_to_fit(nanozone, size, &slot_key); // Note slot_key is set here
    //根据_os_cpu_number经过运算获取 mag_index(meta_data的一维索引)
    mag_index_t mag_index = nano_mag_index(nanozone);
    //确定当前cpu对应的mag和通过size参数计算出来的slot,去对应metadata的链表中取已经被释放过的内存区块缓存
    nano_meta_admin_t pMeta = &(nanozone->meta_data[mag_index][slot_key]);
    //检测是否存在已经释放过,可以直接拿来用的内存,已经被释放的内存会缓存在 chained_block_s 链表
    //每一次free。同样会根据 index 和slot 的值回去 pMeta,然后把slot_LIFO的指针指向释放的内存。
    ptr = OSAtomicDequeue(&(pMeta->slot_LIFO), offsetof(struct chained_block_s, next));
    if (ptr) {
    
    ...省略无关代码
    
    //如果缓存的内存存在,这进行指针地址检查等异常检测,最后返回
    //第一次调用malloc时,不会执行这一块代码。
    } else {
    //没有释放过的内存,所以调用函数 获取内存
        ptr = segregated_next_block(nanozone, pMeta, slot_bytes, mag_index);
    }

    if (cleared_requested && ptr) {
        memset(ptr, 0, slot_bytes); // TODO: Needs a memory barrier after memset to ensure zeroes land first?
    }
    return ptr;
}

该方法主要是通过 cpuslot 确定 index,从chained_block_s 链表中找出是否存在已经释放过的缓存。如果存在则进行指针检查之后返回,否则进入查询 meta data 或者开辟 band

2.9 segregated_next_block 分析

static MALLOC_INLINE void *
segregated_next_block(nanozone_t *nanozone, nano_meta_admin_t pMeta, size_t slot_bytes, unsigned int mag_index)
{
    while (1) {
        //当前这块pMeta可用内存的结束地址
        uintptr_t theLimit = pMeta->slot_limit_addr; // Capture the slot limit that bounds slot_bump_addr right now
        //原子的为pMeta->slot_bump_addr添加slot_bytes的长度,偏移到下一个地址
        uintptr_t b = OSAtomicAdd64Barrier(slot_bytes, (volatile int64_t *)&(pMeta->slot_bump_addr));
        //减去添加的偏移量,获取当前可以获取的地址
        b -= slot_bytes; // Atomic op returned addr of *next* free block. Subtract to get addr for *this* allocation.
        
        if (b < theLimit) {   // Did we stay within the bound of the present slot allocation?
            //如果地址还在范围之内,则返回地址
            return (void *)b; // Yep, so the slot_bump_addr this thread incremented is good to go
        } else {
            //已经用尽了
            if (pMeta->slot_exhausted) { // exhausted all the bands availble for this slot?
                pMeta->slot_bump_addr = theLimit;
                return 0;                 // We're toast
            } else {
                // One thread will grow the heap, others will see its been grown and retry allocation
                _malloc_lock_lock(&nanozone->band_resupply_lock[mag_index]);
                // re-check state now that we've taken the lock
                //多线程的缘故,重新检查是否用尽
                if (pMeta->slot_exhausted) {
                    _malloc_lock_unlock(&nanozone->band_resupply_lock[mag_index]);
                    return 0; // Toast
                } else if (b < pMeta->slot_limit_addr) {
                    //如果小于最大限制地址,当重新申请一个新的band后,重新尝试while
                    _malloc_lock_unlock(&nanozone->band_resupply_lock[mag_index]);
                    continue; // ... the slot was successfully grown by first-taker (not us). Now try again.
                } else if (segregated_band_grow(nanozone, pMeta, slot_bytes, mag_index)) {
                    //申请新的band成功,重新尝试while
                    _malloc_lock_unlock(&nanozone->band_resupply_lock[mag_index]);
                    continue; // ... the slot has been successfully grown by us. Now try again.
                } else {
                    pMeta->slot_exhausted = TRUE;
                    pMeta->slot_bump_addr = theLimit;
                    _malloc_lock_unlock(&nanozone->band_resupply_lock[mag_index]);
                    return 0;
                }
            }
        }
    }
}

如果是第一次调用 segregated_next_block 函数,band 不存在,缓存也不会存在,所以会调用segregated_band_grow。来开辟新的 band

2.10 segregated_band_grow分析

boolean_t
segregated_band_grow(nanozone_t *nanozone, nano_meta_admin_t pMeta, size_t slot_bytes, unsigned int mag_index)
{
    用来计算slot_current_base_addr 的联合体
    nano_blk_addr_t u; // the compiler holds this in a register
    uintptr_t p, s;
    size_t watermark, hiwater;

    if (0 == pMeta->slot_current_base_addr) { // First encounter?
        //利用nano_blk_addr_t 来计算slot_current_base_addr。
        u.fields.nano_signature = NANOZONE_SIGNATURE;
        u.fields.nano_mag_index = mag_index;
        u.fields.nano_band = 0;
        u.fields.nano_slot = (slot_bytes >> SHIFT_NANO_QUANTUM) - 1;
        u.fields.nano_offset = 0;
        
        //根据设置的属性计算 slot_current_base_addr 
        p = u.addr;
        pMeta->slot_bytes = (unsigned int)slot_bytes;
        pMeta->slot_objects = SLOT_IN_BAND_SIZE / slot_bytes;
    } else {
        p = pMeta->slot_current_base_addr + BAND_SIZE; // Growing, so stride ahead by BAND_SIZE

        u.addr = (uint64_t)p;
        if (0 == u.fields.nano_band) { // Did the band index wrap?
            return FALSE;
        }

        assert(slot_bytes == pMeta->slot_bytes);
    }
    pMeta->slot_current_base_addr = p;
//BAND_SIZE = 1 << 21 = 2097152 = 256kb
    mach_vm_address_t vm_addr = p & ~((uintptr_t)(BAND_SIZE - 1)); // Address of the (2MB) band covering this (128KB) slot
    if (nanozone->band_max_mapped_baseaddr[mag_index] < vm_addr) {
    //如果最大能存储的地址 仍然小于目标地址,则小开辟新的band
#if !NANO_PREALLOCATE_BAND_VM
        // Obtain the next band to cover this slot
        //// mac 和模拟器 或重新使用
        // Obtain the next band to cover this slot
        //重新申请新的 band,调用mach_vm_map  从pmap 转换。
        kern_return_t kr = mach_vm_map(mach_task_self(), &vm_addr, BAND_SIZE, 0, VM_MAKE_TAG(VM_MEMORY_MALLOC_NANO),
                MEMORY_OBJECT_NULL, 0, FALSE, VM_PROT_DEFAULT, VM_PROT_ALL, VM_INHERIT_DEFAULT);

        void *q = (void *)vm_addr;
        if (kr || q != (void *)(p & ~((uintptr_t)(BAND_SIZE - 1)))) { // Must get exactly what we asked for
            if (!kr) {
                mach_vm_deallocate(mach_task_self(), vm_addr, BAND_SIZE);
            }
            return FALSE;
        }
#endif
        nanozone->band_max_mapped_baseaddr[mag_index] = vm_addr;
    }

    // Randomize the starting allocation from this slot (introduces 11 to 14 bits of entropy)
    if (0 == pMeta->slot_objects_mapped) { // First encounter?
        pMeta->slot_objects_skipped = (malloc_entropy[1] % (SLOT_IN_BAND_SIZE / slot_bytes));
        pMeta->slot_bump_addr = p + (pMeta->slot_objects_skipped * slot_bytes);
    } else {
        pMeta->slot_bump_addr = p;
    }

    pMeta->slot_limit_addr = p + (SLOT_IN_BAND_SIZE / slot_bytes) * slot_bytes;
    pMeta->slot_objects_mapped += (SLOT_IN_BAND_SIZE / slot_bytes);

    u.fields.nano_signature = NANOZONE_SIGNATURE;
    u.fields.nano_mag_index = mag_index;
    u.fields.nano_band = 0;
    u.fields.nano_slot = 0;
    u.fields.nano_offset = 0;
    s = u.addr; // Base for this core.

    // Set the high water mark for this CPU's entire magazine, if this resupply raised it.
    watermark = nanozone->core_mapped_size[mag_index];
    hiwater = MAX(watermark, p - s + SLOT_IN_BAND_SIZE);
    nanozone->core_mapped_size[mag_index] = hiwater;

    return TRUE;
}

当进入 segregated_band_grow 时,如果当前的 band 不够用,则使用 mach_vm_map 经由 pmap 重新映射物理内存到虚拟内存。

关于通过 nano_blk_addr_t 的联合体结构如下,其每个成员所占的 bit位数 已经写出。

struct nano_blk_addr_s {
    uint64_t                                   
    nano_offset:NANO_OFFSET_BITS,              //17 locates the block
    nano_slot:NANO_SLOT_BITS,                  //4  bucket of homogenous quanta-multiple blocks
    nano_band:NANO_BAND_BITS,                  //17
    nano_mag_index:NANO_MAG_BITS,              //6  the core that allocated this block
    nano_signature:NANOZONE_SIGNATURE_BITS;    //   the address range devoted to us.
};

#endif
// clang-format on

typedef union  {
    uint64_t            addr;
    struct nano_blk_addr_s    fields;
} nano_blk_addr_t;

下面通过 LLDB 分析

image image

free 的阶段,也是使用如上的方式获取 对应的 slot,mag_index

下面来梳理下 nana_zone 分配过程:

  • 确定当前 cpu 对应的 mag 和通过 size参数 计算出来的 slot ,去对应 chained_block_s 的链表中取已经被释放过的内存区块缓存,如果取到检查指针地址是否有问题,没有问题就直接返回;
  • 初次进行 nano malloc时,nano zon并没有缓存,会直接在 nano zone范围的地址空间上直接分配连续地址内存;
  • 如当前 Band 中当前 Slot 耗尽则向系统申请新的 Band(每个 Band固定大小 2M,容纳了16个128k 的槽),连续地址分配内存的基地址、limit地址以及当前分配到的地址由 meta data 结构维护起来,而这些 meta data 则以 MagSlot 为维度(Mag个数是处理器个数,Slot是16个)的二维数组形式,放在 nanozone_tmeta_data字段中。
    流程如下
image

2.11 scalable zone(helper_zone) 分析

szone 上分配的内存包括 tiny、small和large 三大类,其中 tinysmall 的分配、释放过程大致相同,larg类型有自己的方式管理。同样会通过create_scalable_zone来构造zone。 这里不在复述create_scalable_zone`,直接看内存的分配策略

2.12 szone_malloc_should_clear 分析

MALLOC_NOINLINE void *
szone_malloc_should_clear(szone_t *szone, size_t size, boolean_t cleared_requested)
{
    void *ptr;
    msize_t msize;
    //64位 <= 1008B  32位<= 496B
    if (size <= SMALL_THRESHOLD) {
        // tiny size: <=1008 bytes (64-bit), <=496 bytes (32-bit)
        // think tiny
        msize = TINY_MSIZE_FOR_BYTES(size + TINY_QUANTUM - 1);
        if (!msize) {
            msize = 1;
        }
        ptr = tiny_malloc_should_clear(&szone->tiny_rack, msize, cleared_requested);
    } else if (size <= szone->large_threshold) {
        //64位 <= 128KB      32位 <= 128KB
        // small size: <=15k (iOS), <=64k (large iOS), <=128k (macOS)
        // think small
        msize = SMALL_MSIZE_FOR_BYTES(size + SMALL_QUANTUM - 1);
        if (!msize) {
            msize = 1;
        }
        ptr = small_malloc_should_clear(&szone->small_rack, msize, cleared_requested);
    } else {
        // large: all other allocations
        size_t num_kernel_pages = round_page_quanta(size) >> vm_page_quanta_shift;
        if (num_kernel_pages == 0) { /* Overflowed */
            ptr = 0;
        } else {
            ptr = large_malloc(szone, num_kernel_pages, 0, cleared_requested);
        }
    }
#if DEBUG_MALLOC
    if (LOG(szone, ptr)) {
        malloc_report(ASL_LEVEL_INFO, "szone_malloc returned %p\n", ptr);
    }
#endif
    /*
     * If requested, scribble on allocated memory.
     */
    if ((szone->debug_flags & MALLOC_DO_SCRIBBLE) && ptr && !cleared_requested && size) {
        memset(ptr, SCRIBBLE_BYTE, szone_size(szone, ptr));
    }
    return ptr;
}

这里以看出在 szone 上分配的内存包括 tinysmalllarge 三大类,我们以 tiny为例 开始下面的分析

2.12 tiny_malloc_should_clear 分析

void *
tiny_malloc_should_clear(rack_t *rack, msize_t msize, boolean_t cleared_requested)
{
    void *ptr;
    mag_index_t mag_index = tiny_mag_get_thread_index() % rack->num_magazines;
    //获取magazine.  magazines 是一个由64个magazine_t组成的数组
    magazine_t *tiny_mag_ptr = &(rack->magazines[mag_index]);

    MALLOC_TRACE(TRACE_tiny_malloc, (uintptr_t)rack, TINY_BYTES_FOR_MSIZE(msize), (uintptr_t)tiny_mag_ptr, cleared_requested);

#if DEBUG_MALLOC
    if (DEPOT_MAGAZINE_INDEX == mag_index) {
        malloc_zone_error(rack->debug_flags, true, "malloc called for magazine index -1\n");
        return (NULL);
    }

    if (!msize) {
        malloc_zone_error(rack->debug_flags, true, "invariant broken (!msize) in allocation (region)\n");
        return (NULL);
    }
#endif

    SZONE_MAGAZINE_PTR_LOCK(tiny_mag_ptr);

#if CONFIG_TINY_CACHE
    ptr = tiny_mag_ptr->mag_last_free;
    //如果开启了tiny 的缓存。
    if (tiny_mag_ptr->mag_last_free_msize == msize) {
        // we have a winner
        //优先查看上次最后释放的区块是否和此次请求的大小刚好相等(都是对齐之后的slot大小),如果是则直接返回。
        tiny_mag_ptr->mag_last_free = NULL;
        tiny_mag_ptr->mag_last_free_msize = 0;
        tiny_mag_ptr->mag_last_free_rgn = NULL;
        SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr);
        CHECK(szone, __PRETTY_FUNCTION__);
        if (cleared_requested) {
            memset(ptr, 0, TINY_BYTES_FOR_MSIZE(msize));
        }
#if DEBUG_MALLOC
        if (LOG(szone, ptr)) {
            malloc_report(ASL_LEVEL_INFO, "in tiny_malloc_should_clear(), tiny cache ptr=%p, msize=%d\n", ptr, msize);
        }
#endif
        return ptr;
    }
#endif /* CONFIG_TINY_CACHE */

    while (1) {
        //先从freelist 查找
        ptr = tiny_malloc_from_free_list(rack, tiny_mag_ptr, mag_index, msize);
        if (ptr) {
            SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr);
            CHECK(szone, __PRETTY_FUNCTION__);
            if (cleared_requested) {
                memset(ptr, 0, TINY_BYTES_FOR_MSIZE(msize));
            }
            return ptr;
        }
        //从一个后备magazine中取出一个可用region,完整地拿过来放到当前magazine,再走一遍上面的步骤。
        if (tiny_get_region_from_depot(rack, tiny_mag_ptr, mag_index, msize)) {
            //再次尝试从freelist 中获取
            ptr = tiny_malloc_from_free_list(rack, tiny_mag_ptr, mag_index, msize);
            if (ptr) {
                SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr);
                CHECK(szone, __PRETTY_FUNCTION__);
                if (cleared_requested) {
                    memset(ptr, 0, TINY_BYTES_FOR_MSIZE(msize));
                }
                return ptr;
            }
        }

        // The magazine is exhausted. A new region (heap) must be allocated to satisfy this call to malloc().
        // The allocation, an mmap() system call, will be performed outside the magazine spin locks by the first
        // thread that suffers the exhaustion. That thread sets "alloc_underway" and enters a critical section.
        // Threads arriving here later are excluded from the critical section, yield the CPU, and then retry the
        // allocation. After some time the magazine is resupplied, the original thread leaves with its allocation,
        // and retry-ing threads succeed in the code just above.
        if (!tiny_mag_ptr->alloc_underway) {
            //如果没有正在申请新的的 regin 操作,则进行申请操作
            void *fresh_region;
            
            // time to create a new region (do this outside the magazine lock)
            //设置当前正在申请新的 堆
            tiny_mag_ptr->alloc_underway = TRUE;
            OSMemoryBarrier();
            SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr);
            //申请新的堆                                 1m             
            fresh_region = mvm_allocate_pages_securely(TINY_REGION_SIZE, TINY_BLOCKS_ALIGN, VM_MEMORY_MALLOC_TINY, rack->debug_flags);
            SZONE_MAGAZINE_PTR_LOCK(tiny_mag_ptr);

            // DTrace USDT Probe
            MAGMALLOC_ALLOCREGION(TINY_SZONE_FROM_RACK(rack), (int)mag_index, fresh_region, TINY_REGION_SIZE);

            if (!fresh_region) { // out of memory!
                tiny_mag_ptr->alloc_underway = FALSE;
                OSMemoryBarrier();
                SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr);
                return NULL;
            }
            //从最近的一个 region 或者新申请的 region中malloc
            ptr = tiny_malloc_from_region_no_lock(rack, tiny_mag_ptr, mag_index, msize, fresh_region);

            // we don't clear because this freshly allocated space is pristine
            tiny_mag_ptr->alloc_underway = FALSE;
            OSMemoryBarrier();
            SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr);
            CHECK(szone, __PRETTY_FUNCTION__);
            return ptr;
        } else {
            SZONE_MAGAZINE_PTR_UNLOCK(tiny_mag_ptr);
            yield();
            SZONE_MAGAZINE_PTR_LOCK(tiny_mag_ptr);
        }
    }
    /* NOTREACHED */
}

每次调用 free 函数,会直接把要释放的内存优先放到mag_last_free 指针上,在下次 alloc 时,也会优先检查mag_last_free 是否存在大小相等的内存,如果存在就直接返回。

2.14 tiny_malloc_from_free_list & tiny_get_region_from_depot 分析

每一个类型的 rack 指向的 magazines ,都会在下标为-1 , magazine_t 当做备用:depot,该方法的作用是从备用的 depot查找出是否有满足条件的 region 如果存在,更新 depotregion 的关联关系,然后在关联当前的magazine_tregion。之后在再次重复 free_list 过程

2.15 mvm_allocate_pages_securely 的分析

image

2.16 tiny_malloc_from_region_no_lock 的分析

重新申请了新的内存 (region) 之后,挂载到当前的 magazine下并分配内存。

这个方法的主要作用是把新申请的内存地址,转换为region,并进行相关的关联。及更新对应的 magazine。整个 scalable_zone 的结构体关系,及流程如下

image

2.17 nano_zone 总结

malloc 库会检查指针地址,如果没有问题,则以链表的形式将这些区块按大小存储起来。这些链表的头部放在 meta_data数组 中对应的 [mag][slot]元素中。

其实从缓存获取空余内存和释放内存时都会对指向这篇内存区域的指针进行检查,如果有类似地址不对齐、未释放/多次释放、所属地址与预期的 mag、slot 不匹配等情况都会以报错结束。

2.18 scalable_zone 分析

三、流程总结

image

四、拓展补充

上一篇 下一篇

猜你喜欢

热点阅读