作者:eddiecmchen,PCG客户端开发工程师
| 导语 把我的iphone XR扶起来,它还能再顶一会儿~
远在IOS 11时期(2017年),苹果就发公告要求所有需要上架AppStore的应用都必须支持64位。32位应用不再支持上架与运行。
升级64位应用有什么好处呢?(以下内容纯摘抄,客官可以直接跳过)
这里我们要注意的是:虚拟内存确实比纯32位多了,但是App到底能用多少,是否跟宣传一样接近16EB?下面将会展开聊聊,我们先来看一个Crash。
我们先来看下面的一个内存导致的崩溃,JSC在使用bmalloc尝试进行内存分配时,提示OOM导致了SIGTRAP。
Last Exception :
0 JAVAScriptCore 0x000000018b777570 _pas_panic_on_out_of_memory_error
1 JavaScriptCore 0x000000018b72e918 _bmalloc_try_iso_allocate_impl_impl_slow
2 JavaScriptCore 0x000000018b73d3d8 _bmalloc_heap_config_specialized_local_allocator_try_allocate_small_segregated_slow + 5952
3 JavaScriptCore 0x000000018b7276f8 _bmalloc_allocate_impl_casual_case + 800
4 JavaScriptCore 0x000000018c60d494 JSC::PropertyTable::create(JSC::VM&, unsigned int) + 244
5 JavaScriptCore 0x000000018c66ba74 JSC::Structure::materializePropertyTable(JSC::VM&, bool) + 324
6 JavaScriptCore 0x000000018c66dfac JSC::Structure::changePrototypeTransition(JSC::VM&, JSC::Structure*, JSC::JSValue, JSC::DeferredStructureTransitionWatchpointFire&) + 612
7 JavaScriptCore 0x000000018c559930 JSC::JSObject::setPrototypeDirect(JSC::VM&, JSC::JSValue) + 192
8 JavaScriptCore 0x000000018c559e40 JSC::JSObject::setPrototypeWithCycleCheck(JSC::VM&, JSC::JSGlobalObject*, JSC::JSValue, bool) + 316
9 JavaScriptCore 0x000000018c4f580c JSC::globalFuncProtoSetter(JSC::JSGlobalObject*, JSC::CallFrame*) + 192
10 JavaScriptCore 0x000000018ba1f7a8 _vmEntryToNative + 280
11 JavaScriptCore 0x000000018c1b0cd0 JSC::Interpreter::executeCall(JSC::JSGlobalObject*, JSC::JSObject*, JSC::CallData const&, JSC::JSValue, JSC::ArgList const&) + 616
12 JavaScriptCore 0x000000018c474ecc JSC::GetterSetter::callSetter(JSC::JSGlobalObject*, JSC::JSValue, JSC::JSValue, bool) + 212
13 JavaScriptCore 0x000000018c5b6264 JSC::JSGenericTypedArrayView<JSC::Uint8Adaptor>::put(JSC::JSCell*, JSC::JSGlobalObject*, JSC::PropertyName, JSC::JSValue, JSC::PutPropertySlot&) + 612
14 JavaScriptCore 0x000000018c2c2ecc _llint_slow_path_put_by_id + 3244
// 忽略多余重复堆栈
37 JavaScriptCore 0x000000018ba1f5fc _vmEntryToJavaScript + 264
38 JavaScriptCore 0x000000018c1b0c7c JSC::Interpreter::executeCall(JSC::JSGlobalObject*, JSC::JSObject*, JSC::CallData const&, JSC::JSValue, JSC::ArgList const&) + 532
39 JavaScriptCore 0x000000018bac7ae4 _JSObjectCallAsFunction + 568
40 mttlite 0x0000000102a54914 hippy::napi::JSCCtx::CallFunction(std::__1::shared_ptr<hippy::napi::CtxValue> const&, unsigned long, std::__1::shared_ptr<hippy::napi::CtxValue> const*) (js_native_api_value_jsc.cc:406)
41 mttlite 0x0000000102a664e0 _ZNSt3__110__function6__funcIZN11TimerModule5StartERKN5hippy4napi12CallbackInfoEbE3$_4NS_9allocatorIS8_EEFvvEEclEv (memory:3237)
42 mttlite 0x0000000102a63018 hippy::base::TaskRunner::Run() (memory:3237)
43 mttlite 0x0000000102a64974 ThreadEntry (thread.cc:0)
44 libsystem_pthread.dylib 0x00000001dc129348 __pthread_start + 116
------
Exception Type: SIGTRAP
Exception Codes: fault addr: 0x000000018b777570
Crashed Thread: 48 hippy.js
这个OOM问题,与iOS上常见的OOM不一样。按照常规的理解,当App内存不足的时候,正常会触发系统的Jetsam机制杀死App。在系统日志中会留下Jetsam相关日志,理论上不会在Bugly等异常上报中发现。但这一类崩溃却一直在产生上报,并且低内存的崩溃堆栈表现形式有很多种。
以上的JSC崩溃问题已经存在很长一段时间了(至少2年),而且崩溃堆栈都集中在JSC执行JS代码的过程中,长期缺乏JS相关的监控与Debug工具导致该问题一直无法解决。
虽然堆栈上有明确的原因说明是OOM,但我们观察到有不少用户实际上物理内存空间还是足够的:
两年前,冲浪的时候偶然看来了来自微视同学的Case总结:《OOM与内存》
当时跟hippy SDK的同事也讨论过是否存在类似的内存不足情况。但由于大家对JSC黑盒都不熟悉,而且崩溃的JS堆栈也不确切。当时的建议是:少在后台加载JSC。最终也并没有解决该问题。
两年后,当浏览器集成flutter,类似的JS崩溃直接翻倍(21H2 0.08% -> 22H1 0.16%)。没办法,还是要看类似JSC和Dart VM的内存分配机制是怎样的,再挖掘一下是否存在解(缓)决(解)方案。
翻阅相关虚拟机的内存管理相关代码,可以找到底层的内存分配基本实现都是基于mmap处理的。
// WebKit bmalloc VMAllocate
inline void* tryVMAllocate(size_t vmSize, VMTag usage = VMTag::Malloc)
{
vmValidate(vmSize);
void* result = mmap(0, vmSize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON | BMALLOC_NORESERVE, static_cast<int>(usage), 0);
if (result == MAP_FAILED)
return nullptr;
return result;
}
// Dart VM的虚拟内存
VirtualMemory* VirtualMemory::Allocate(intptr_t size,
bool is_executable,
const char* name) {
ASSERT(Utils::IsAligned(size, PageSize()));
const int prot = PROT_READ | PROT_WRITE | (is_executable ? PROT_EXEC : 0);
int map_flags = MAP_PRIVATE | MAP_ANONYMOUS;
#if (defined(DART_HOST_OS_macOS) && !defined(DART_HOST_OS_IOS))
if (is_executable && IsAtLeastOS10_14()) {
map_flags |= MAP_JIT;
}
#endif // defined(DART_HOST_OS_MACOS)
// Some 64-bit microarchitectures store only the low 32-bits of targets as
// part of indirect branch prediction, predicting that the target's upper bits
// will be same as the call instruction's address. This leads to misprediction
// for indirect calls crossing a 4GB boundary. We ask mmap to place our
// generated code near the VM binary to avoid this.
void* hint = is_executable ? reinterpret_cast<void*>(&Allocate) : nullptr;
void* address = mmap(hint, size, prot, map_flags, -1, 0);
if (address == MAP_FAILED) {
return nullptr;
}
return new VirtualMemory(address, size);
}
VirtualMemory::~VirtualMemory() {
if (address_ != nullptr) {
if (munmap(address_, size_) != 0) {
int error = errno;
const int kBufferSize = 1024;
char error_buf[kBufferSize];
FATAL("munmap error: %d (%s)", error,
Utils::StrError(error, error_buf, kBufferSize));
}
}
}
当map_flags包含MAP_ANON时,并且fd传入-1时,mmap将直接使用虚拟内存进行分配,不需要依赖文件描述符。
/*
* mmap stub, with preemptory failures due to extra parameter checking
* mandated for conformance.
*
* This is for UNIX03 only.
*/
void *
mmap(void *addr, size_t len, int prot, int flags, int fildes, off_t off)
{
/*
* Preemptory failures:
*
* o off is not a multiple of the page size
* o flags does not contain either MAP_PRIVATE or MAP_SHARED
* o len is zero
*/
extern void cerror_nocancel(int);
if ((off & PAGE_MASK) ||
(((flags & MAP_PRIVATE) != MAP_PRIVATE) &&
((flags & MAP_SHARED) != MAP_SHARED)) ||
(len == 0)) {
cerror_nocancel(EINVAL);
return(MAP_FAILED);
}
void *ptr = __mmap(addr, len, prot, flags, fildes, off);
if (__syscall_logger) {
int stackLoggingFlags = stack_logging_type_vm_allocate;
if (flags & MAP_ANON) {
stackLoggingFlags |= (fildes & VM_FLAGS_ALIAS_MASK);
} else {
stackLoggingFlags |= stack_logging_type_mapped_file_or_shared_mem;
}
__syscall_logger(stackLoggingFlags, (uintptr_t)mach_task_self(), (uintptr_t)len, 0, (uintptr_t)ptr, 0);
}
return ptr;
}
上面的调用会传递到内核kern_mman.c的实现函数mmap(proc_t p, struct mmap_args *uap, user_addr_t *retval)
/*
* XXX Internally, we use VM_PROT_* somewhat interchangeably, but the correct
* XXX usage is PROT_* from an interface perspective. Thus the values of
* XXX VM_PROT_* and PROT_* need to correspond.
*/
int
mmap(proc_t p, struct mmap_args *uap, user_addr_t *retval)
{
/*
* 上面忽略了一部分代码
*/
result = vm_map_enter_mem_object(user_map,
&user_addr, user_size,
0, alloc_flags, vmk_flags,
tag,
IPC_PORT_NULL, 0, FALSE,
prot, maxprot,
(flags & MAP_SHARED) ?
VM_INHERIT_SHARE :
VM_INHERIT_DEFAULT);
/* If a non-binding address was specified for this anonymous
* mapping, retry the mapping with a zero base
* in the event the mapping operation failed due to
* lack of space between the address and the map's maximum.
*/
if ((result == KERN_NO_SPACE) && ((flags & MAP_FIXED) == 0) && user_addr && (num_retries++ == 0)) {
user_addr = vm_map_page_size(user_map);
goto map_anon_retry;
}
/*
* 下面忽略了一部分代码
*/
}
其中又会调用vm_map.c内部的vm_map_enter_mem_object,而该方法最终会在vm_map_enter中依据对象进行内存分配:
// 下面这个只截了个头,大概带一下,我也没调过代码~
/*
* Routine: vm_map_enter
*
* Description:
* Allocate a range in the specified virtual address map.
* The resulting range will refer to memory defined by
* the given memory object and offset into that object.
*
* Arguments are as defined in the vm_map call.
*/
kern_return_t
vm_map_enter(
vm_map_t map,
vm_map_offset_t *address, /* IN/OUT */
vm_map_size_t size,
vm_map_offset_t mask,
int flags,
vm_map_kernel_flags_t vmk_flags,
vm_tag_t alias,
vm_object_t object,
vm_object_offset_t offset,
boolean_t needs_copy,
vm_prot_t cur_protection,
vm_prot_t max_protection,
vm_inherit_t inheritance)
其中vm_map_enter在分配过程中会对hole_entry→vme_end作判断,vme_end即最大的可分配空间。
本来我只是观察到苹果在iOS15上增加了com.apple.developer.kernel.increased-memory-limit的能力声明。本着死马当活马医的想法,尝试在新版本上添加该声明以缓解一部分问题。
结果偶然看到部分开发者提问:该能力可配合com.apple.developer.kernel.extended-virtual-addressing使用。看到后我一下子反应过来,顺手搜到了今年二月国外有大佬做了相关的探索:
Size Matters: An Exploration of Virtual Memory on iOS
文章阐述了iOS的内存管理机制和虚拟内存空间分配在不同的机型上存在上限,代码如下:
#define ARM64_MIN_MAX_ADDRESS (SHARED_REGION_BASE_ARM64 + SHARED_REGION_SIZE_ARM64 + 0x20000000) // end of shared region + 512MB for various purposes
const vm_map_offset_t min_max_offset = ARM64_MIN_MAX_ADDRESS; // end of shared region + 512MB for various purposes
if (arm64_pmap_max_offset_default) {
max_offset_ret = arm64_pmap_max_offset_default;
} else if (max_mem > 0xC0000000) {
max_offset_ret = min_max_offset + 0x138000000; // Max offset is 13.375GB for devices with > 3GB of memory
} else if (max_mem > 0x40000000) {
max_offset_ret = min_max_offset + 0x38000000; // Max offset is 9.375GB for devices with > 1GB and <= 3GB of memory
} else {
max_offset_ret = min_max_offset;
}
并且总结了一个上限值与机型表格:
RAM |
Address Space |
Usable |
Devices |
> 3 GiB |
15.375 GiB |
7.375 GiB |
- iPhone XS – iPhone 13 |
> 1 GiB |
11.375 GiB |
3.375 GiB |
- iPhone 6s – X, SE, XR |
<= 1 GiB |
10.5 GiB |
2.5 GiB |
- iPhone 5s, iPhone 6 |
而xnu的源码(pmap.c)中还透露了内核内存分配存在jumbo机制。当iOS App带有指定的能力声明时,xnu内核将会以jumbo模式运行,虚拟内存地址空间将会直接分配为最大值64GB:
if (option == ARM_PMAP_MAX_OFFSET_JUMBO) {
if (arm64_pmap_max_offset_default) {
// Allow the boot-arg to override jumbo size
max_offset_ret = arm64_pmap_max_offset_default;
} else {
max_offset_ret = MACH_VM_MAX_ADDRESS; // Max offset is 64GB for pmaps with special "jumbo" blessing
}
}
并且该上限值会在进程启动时进行调整,具体代码可以在kern_exec.c中找到:
/*
* Apply the requested maximum address.
*/
if (error == 0 && imgp->ip_px_sa != NULL) {
struct _posix_spawnattr *psa = (struct _posix_spawnattr *) imgp->ip_px_sa;
if (psa->psa_max_addr) {
vm_map_set_max_addr(get_task_map(new_task), (vm_map_offset_t)psa->psa_max_addr);
}
}
com.apple.developer.kernel.extended-virtual-addressing
苹果的文档仅有一句话说明该能力:
Use this entitlement if your app has specific needs that require a larger addressable space. For example, games that memory map assets to stream to the GPU may benefit from a larger address space.
举个例子:有的游戏需要将资源通过mmap的形式传递到GPU中渲染时,更大的地址空间可提高其运行效率。
描述上看,配置该选项时,将开启上面xnu的jumbo mode,地址的扩充刚好能解决上面的崩溃问题。
为验证地址分配的极限值,简单做个实验(测试设备使用iPhone XR iOS 16 Beta 2):
通过malloc进行连续的内存分配(也可以用vm_allocate,阈值不一样),阈值卡在1009字节(为什么是1009字节,这里可以参考【ios 内核】源码解读(3) 详解ios是怎么malloc的(上) - 钟路成的博客 (luchengzhong.github.io))。
for (size_t i = 0; i < SIZE_T_MAX; i++) {
void *a = malloc(1009);
if (a == NULL) {
NSLog(@"error count: %lu", i);
break;
}
}
结果如下:
size = 1009 > SMALL_THRESHOLD (64位系统下1008字节,32位系统下496)
内存扩展前malloc失败阈值约 7065482 * 1009 = 6.63 GB
内存扩展后malloc失败阈值约 56753881 * 1009 = 53.33 GB
当然,在xnu的单元测试代码中,也可找到jumbo mode相关的测试代码,与上面的测试结果完全一致,即最多可分配53GB的空间。
#define GB (1ULL * 1024 * 1024 * 1024)
/*
* This test expects the entitlement to be the enabling factor for a process to
* allocate at least this many GB of VA space. i.e. with the entitlement, n GB
* must be allocatable; whereas without it, it must be less.
* This value was determined experimentally to fit on applicable devices and to
* be clearly distinguishable from the default VA limit.
*/
#define ALLOC_TEST_GB 53
T_DECL(TESTNAME,
"Verify that a required entitlement is present in order to be granted an extra-large "
"VA space on arm64",
T_META_NAMESPACE("xnu.vm"),
T_META_CHECK_LEAKS(false))
{
int i;
void *res;
if (!dt_64_bit_kernel()) {
T_SKIP("This test is only applicable to arm64");
}
T_LOG("Attemping to allocate VA space in 1 GB chunks.");
for (i = 0; i < (ALLOC_TEST_GB * 2); i++) {
res = mmap(NULL, 1 * GB, PROT_NONE, MAP_PRIVATE | MAP_ANON, 0, 0);
if (res == MAP_FAILED) {
if (errno != ENOMEM) {
T_WITH_ERRNO;
T_LOG("mmap failed: stopped at %d of %d GB allocated", i, ALLOC_TEST_GB);
}
break;
} else {
T_LOG("%d: %pn", i, res);
}
}
#if defined(ENTITLED)
T_EXPECT_GE_INT(i, ALLOC_TEST_GB, "Allocate at least %d GB of VA space", ALLOC_TEST_GB);
#else
T_EXPECT_LT_INT(i, ALLOC_TEST_GB, "Not permitted to allocate %d GB of VA space", ALLOC_TEST_GB);
#endif
}
可见,当开启com.apple.developer.kernel.extended-virtual-addressing时,内核的可分配空间确实有明显提升。
从QQ浏览器的上线效果来看,JS相关的内存分配Crash在14.0以上系统几乎全部消失。上线第一天App崩溃率环比下降接近50%,效果显著。
简单总结:
以上源码相关的内容仅个人阅读理解,如有错误请指出。