99% iOS开发都不知道的KVO崩溃分析详解
背景
crash 监控发现有大量的新增崩溃,堆栈如下
libsystem_platform.dylib __os_unfair_lock_corruption_abort() libsystem_platform.dylib __os_unfair_lock_lock_slow() Foundation __NSSetBoolValueAndNotify()
分析堆栈
__os_unfair_lock_corruption_abort
log 翻译:lock 已损坏
_os_unfair_lock_corruption_abort(os_ulock_value_t current) { __LIBPLATFORM_CLIENT_CRASH__(current, "os_unfair_lock is corrupt"); }
__os_unfair_lock_lock_slow
在这个方法里面 __ulock_wait 返回 EOWNERDEAD 调用 corruption abort 方法。
int ret = __ulock_wait(UL_UNFAIR_LOCK | ULF_NO_ERRNO | options, l, current, 0); if (unlikely(ret < 0)) { switch (-ret) { case EINTR: case EFAULT: continue; case EOWNERDEAD: _os_unfair_lock_corruption_abort(current); break; default: __LIBPLATFORM_INTERNAL_CRASH__(-ret, "ulock_wait failure"); } }
EOWNERDEAD 的定义
#define EOWNERDEAD 105 /* Previous owner died */
到这里猜测是 lock 的 owner 已经野指针了,继续向下看。
__NSSetBoolValueAndNotify
google 下这个方法是在 KVO 里面修改属性的时候调用,伪代码:
int __NSSetBoolValueAndNotify(int arg0, int arg1, int arg2) { r31 = r31 - 0x90; var_30 = r24; stack[-56] = r23; var_20 = r22; stack[-40] = r21; var_10 = r20; stack[-24] = r19; saved_fp = r29; stack[-8] = r30; r20 = arg2; r21 = arg1; r19 = arg0; r0 = object_getClass(arg0); r0 = object_getIndexedIvars(r0); // 理清这个崩溃的关键方法,这里和汇编代码不一致,汇编代码的入参是 r0 + 0x20 r23 = r0; os_unfair_recursive_lock_lock_with_options(); CFDictionaryGetValue(*(r23 + 0x18), r21); r22 = _objc_msgSend$copyWithZone:(); os_unfair_recursive_lock_unlock(); if (*(int8_t *)(r23 + 0x28) != 0x0) { _objc_msgSend$willChangeValueForKey:(); (class_getMethodImplementation(*r23, r21))(r19, r21, r20); _objc_msgSend$didChangeValueForKey:(); } else { _objc_msgSend$_changeValueForKey:key:key:usingBlock:(); } var_38 = **qword_9590e8; r0 = objc_release_x22(); if (**qword_9590e8 != var_38) { r0 = __stack_chk_fail(); } return r0; }
os_unfair_recursive_lock_lock_with_options
崩溃调用栈中间还有这一层的内联调用 os_unfair_recursive_lock_lock_with_options。这里的 lock owner 有个比较赋值的操作,如果 oul_value 等于 OS_LOCK_NO_OWNER 则赋值 self 然后 return。崩溃时这里继续向下执行了,那这里的 oul_value 的取值只能是 lock->oul_value。到这里猜测崩溃的原因是 lock->oul_value 野指针了。
void os_unfair_recursive_lock_lock_with_options(os_unfair_recursive_lock_t lock, os_unfair_lock_options_t options) { os_lock_owner_t cur, self = _os_lock_owner_get_self(); _os_unfair_lock_t l = (_os_unfair_lock_t)&lock->ourl_lock; if (likely(os_atomic_cmpxchgv2o(l, oul_value, OS_LOCK_NO_OWNER, self, &cur, acquire))) { return; } if (OS_ULOCK_OWNER(cur) == self) { lock->ourl_count++; return; } return _os_unfair_lock_lock_slow(l, self, options); } OS_ALWAYS_INLINE OS_CONST static inline os_lock_owner_t _os_lock_owner_get_self(void) { os_lock_owner_t self; self = (os_lock_owner_t)_os_tsd_get_direct(__TSD_MACH_THREAD_SELF); return self; }
object_getIndexedIvars
__NSSetBoolValueAndNotify 里面的获取 lock 的方法,这个函数非常关键。
/** * Returns a pointer to any extra bytes allocated with an instance given object. * * @param obj An Objective-C object. * * @return A pointer to any extra bytes allocated with \e obj. If \e obj was * not allocated with any extra bytes, then dereferencing the returned pointer is undefined. * * @note This function returns a pointer to any extra bytes allocated with the instance * (as specified by \c class_createInstance with extraBytes>0). This memory follows the * object's ordinary ivars, but may not be adjacent to the last ivar. * @note The returned pointer is guaranteed to be pointer-size aligned, even if the area following * the object's last ivar is less aligned than that. Alignment greater than pointer-size is never * guaranteed, even if the area following the object's last ivar is more aligned than that. * @note In a garbage-collected environment, the memory is scanned conservatively. /** * Returns a pointer immediately after the instance variables declared in an * object. This is a pointer to the storage specified with the extraBytes * parameter given when allocating an object. */ void *object_getIndexedIvars(id obj) { uint8_t *base = (uint8_t *)obj; if (_objc_isTaggedPointerOrNil(obj)) return nil; if (!obj->isClass()) return base + obj->ISA()->alignedInstanceSize(); Class cls = (Class)obj; if (!cls->isAnySwift()) return base + sizeof(objc_class); swift_class_t *swcls = (swift_class_t *)cls; return base - swcls->classAddressOffset + word_align(swcls->classSize); }
上层调用 __NSSetBoolValueAndNotify 里面:
r0 = object_getClass(arg0),arg0 是实例对象,r0 是类对象,因为这里是个 KVO 的调用,那正常情况下r0 是 NSKVONotifying_xxx。
对于 KVO 类,object_getIndexedIvars 返回的地址是 (uint8_t *)obj + sizeof(objc_class)。根据函数的注释,这个地址指向创建类时附在类空间后 extraBytes 大小的一块内存。
debug 调试
object_getIndexedIvars
__NSSetBoolValueAndNotify 下的调用
object_getIndexedIvars 入参是 NSKVONotifying_KVObject,object_getClass 获取的是 KVO Class。
objc_allocateClassPair
动态创建 KVO 类的方法。
thread #8, queue = 'com.apple.root.default-qos', stop reason = breakpoint 1.1 * frame #0: 0x000000018143a088 libobjc.A.dylib`objc_allocateClassPair frame #1: 0x000000018259cd94 Foundation`_NSKVONotifyingCreateInfoWithOriginalClass + 152 frame #2: 0x00000001825b8fd0 Foundation`_NSKeyValueContainerClassGetNotifyingInfo + 56 frame #3: 0x000000018254b7dc Foundation`-[NSKeyValueUnnestedProperty _isaForAutonotifying] + 44 frame #4: 0x000000018254b504 Foundation`-[NSKeyValueUnnestedProperty isaForAutonotifying] + 88 frame #5: 0x000000018254b32c Foundation`-[NSObject(NSKeyValueObserverRegistration) _addObserver:forProperty:options:context:] + 404 frame #6: 0x000000018254b054 Foundation`-[NSObject(NSKeyValueObserverRegistration) addObserver:forKeyPath:options:context:] + 136 frame #7: 0x00000001040d1860 Test`__29-[ViewController viewDidLoad]_block_invoke(.block_descriptor=0x0000000282a55170) at ViewController.m:28:13 frame #8: 0x00000001043d05a8 libdispatch.dylib`_dispatch_call_block_and_release + 32 frame #9: 0x00000001043d205c libdispatch.dylib`_dispatch_client_callout + 20 frame #10: 0x00000001043d4b94 libdispatch.dylib`_dispatch_queue_override_invoke + 1052 frame #11: 0x00000001043e6478 libdispatch.dylib`_dispatch_root_queue_drain + 408 frame #12: 0x00000001043e6e74 libdispatch.dylib`_dispatch_worker_thread2 + 196 frame #13: 0x00000001d515fdbc libsystem_pthread.dylib`_pthread_wqthread + 228
_NSKVONotifyingCreateInfoWithOriginalClass
objc_allocateClassPair 的上层调用。 allocate 之前的 context w2 是个固定值 0x30,即创建 KVO Class 入参 extraBytes 的大小是 0x30
0x18259cd78 <+124>: mov x1, x21 0x18259cd7c <+128>: mov x2, x22 0x18259cd80 <+132>: bl 0x188097080 0x18259cd84 <+136>: mov x0, x20 0x18259cd88 <+140>: mov x1, x19 0x18259cd8c <+144>: mov w2, #0x30 0x18259cd90 <+148>: bl 0x1880961f0 // objc_allocateClassPair 0x18259cd94 <+152>: cbz x0, 0x18259ce24 ; <+296> 0x18259cd98 <+156>: mov x21, x0 0x18259cd9c <+160>: bl 0x188096410 // objc_registerClassPair 0x18259cda0 <+164>: mov x0, x19 0x18259cda4 <+168>: bl 0x182b45f44 ; symbol stub for: free 0x18259cda8 <+172>: mov x0, x21 0x18259cdac <+176>: bl 0x1880967e0 // object_getIndexedIvars 0x18259cdb0 <+180>: mov x19, x0 0x18259cdb4 <+184>: stp x20, x21, [x0]
_NSKVONotifyingCreateInfoWithOriginalClass+184 处将 x20 和 x21 写入 [x0],此时 x0 指向的是大小为 extraBytes 的内存,打印 x20 和 x21 的值
x20 = 0x00000001117caa10 (void *)0x00000001117caa38: KVObject(向上回溯这个值取自 _NSKVONotifyingCreateInfoWithOriginalClass 的入参 x0)
x21 NSKVONotifying_KVObject
根据这里可以看出 object_getIndexedIvars 返回的地址,依次存储了 KVObject(origin Class) 和 NSKVONotifying_KVObject(KVO Class)。
查看 _NSKVONotifyingCreateInfoWithOriginalClass 的伪代码,对 [x0] 有 5 次写入的操作,并且最终这个方法返回的是 x0 的地址。
function __NSKVONotifyingCreateInfoWithOriginalClass { r31 = r31 - 0x50; stack[32] = r22; stack[40] = r21; stack[48] = r20; stack[56] = r19; stack[64] = r29; stack[72] = r30; r20 = r0; if (*(int8_t *)0x993e78 != 0x0) { os_unfair_lock_assert_owner(0x993e7c); } r0 = class_getName(r20); r22 = strlen(r0) + 0x10; r0 = malloc(r22); r19 = r0; strlcpy(r0, "NSKVONotifying_", r22); strlcat(r19, r21, r22); r0 = objc_allocateClassPair(r20, r19, 0x30); if (r0 != 0x0) { objc_registerClassPair(r0); free(r19); r0 = object_getIndexedIvars(r21); r19 = r0; *(int128_t *)r0 = r20; // 第一次写入 Class *(int128_t *)(r0 + 0x8) = r21; // 第二次写入 Class *(r19 + 0x10) = CFSetCreateMutable(0x0, 0x0, *qword_9592d8); // 第三次写入 CFSet *(int128_t *)(r19 + 0x18) = CFDictionaryCreateMutable(0x0, 0x0, 0x0, *qword_959598); // 第四次写入 CFDictionary *(int128_t *)(r19 + 0x20) = 0x0; // 第五次写入空值 if (*qword_9fc560 != -0x1) { dispatch_once(0x9fc560, 0x8eaf98); } if (class_getMethodImplementation(*r19, @selector(willChangeValueForKey:)) != *qword_9fc568) { r8 = 0x1; } else { r0 = *r19; r0 = class_getMethodImplementation(r0, @selector(didChangeValueForKey:)); r8 = *qword_9fc570; if (r0 != r8) { r8 = *qword_9fc570; if (CPU_FLAGS & NE) { r8 = 0x1; } } } *(int8_t *)(r19 + 0x28) = r8; _NSKVONotifyingSetMethodImplementation(r19, @selector(_isKVOA), 0x44fab4, 0x0); _NSKVONotifyingSetMethodImplementation(r19, @selector(dealloc), 0x44fabc, 0x0); _NSKVONotifyingSetMethodImplementation(r19, @selector(class), 0x44fd2c, 0x0); } else { if (*qword_9fc558 != -0x1) { dispatch_once(0x9fc558, 0x8eaf78); } if (os_log_type_enabled(*0x9fc550, 0x10) != 0x0) { _os_log_error_impl(0x0, *0x9fc550, 0x10, "KVO failed to allocate class pair for name %s, automatic key-value observing will not work for this class", &stack[0], 0xc); } free(r19); r19 = 0x0; } if (**qword_9590e8 == **qword_9590e8) { r0 = r19; } else { r0 = __stack_chk_fail(); } return r0; }
_NSKVONotifyingCreateInfoWithOriginalClass 的上层调用,入参是 [x19, #0x8],返回的参数写入 [x19, #0x28]
0x1825b8fc0 <+40>: ldr x0, [x19, #0x28] 0x1825b8fc4 <+44>: b 0x1825b8fd4 ; <+60> 0x1825b8fc8 <+48>: ldr x0, [x19, #0x8] -> 0x1825b8fcc <+52>: bl 0x18259ccfc ; _NSKVONotifyingCreateInfoWithOriginalClass 0x1825b8fd0 <+56>: str x0, [x19, #0x28] 0x1825b8fd4 <+60>: ldp x29, x30, [sp, #0x10] 0x1825b8fd8 <+64>: ldp x20, x19, [sp], #0x20
打印 x19 是一个 NSKeyValueContainerClass 类型的实例对象,这个对象类的 ivars layout
ivars 0x99f3c0 __OBJC_$_INSTANCE_VARIABLES_NSKeyValueContainerClass entsize 32 count 5 offset 0x9e6048 _OBJC_IVAR_$_NSKeyValueContainerClass._originalClass 8 name 0x90bd27 _originalClass type 0x929ae6 # alignment 3 size 8 offset 0x9e6050 _OBJC_IVAR_$_NSKeyValueContainerClass._cachedObservationInfoImplementation 16 name 0x90bd36 _cachedObservationInfoImplementation type 0x92bb88 ^? alignment 3 size 8 offset 0x9e6058 _OBJC_IVAR_$_NSKeyValueContainerClass._cachedSetObservationInfoImplementation 24 name 0x90bd5b _cachedSetObservationInfoImplementation type 0x92bb88 ^? alignment 3 size 8 offset 0x9e6060 _OBJC_IVAR_$_NSKeyValueContainerClass._cachedSetObservationInfoTakesAnObject 32 name 0x90bd83 _cachedSetObservationInfoTakesAnObject type 0x92a01a B alignment 0 size 1 offset 0x9e6068 _OBJC_IVAR_$_NSKeyValueContainerClass._notifyingInfo 40 name 0x90bdaa _notifyingInfo type 0x92bdd7 ^{?=##^{__CFSet}^{__CFDictionary}{os_unfair_recursive_lock_s={os_unfair_lock_s=I}I}B} alignment 3 size 8
offset 0x8 name:_originalClass type:Class
offset 0x28 name:_notifyingInfo type:struct
_notifyingInfo 结构体
{ Class, Class, __CFSet, __CFDictionary, os_unfair_recursive_lock_s }
type encoding:
developer.apple.com/library/arc…
从 context 可以看出_NSKVONotifyingCreateInfoWithOriginalClass 这个方法入参是 OBJC_IVAR_NSKeyValueContainerClass._originalClass。
返回值 x0 是 _OBJC_IVAR__NSKeyValueContainerClass._notifyingInfo。5 次对 [x0] 的写入是在初始化 _notifyingInfo。
崩溃时的 context:
0x1825231f0 <+56>: bl 0x1880967c0 // object_getClass 0x1825231f4 <+60>: bl 0x1880967e0 // object_getIndexedIvars 0x1825231f8 <+64>: mov x23, x0 // x0 == _notifyingInfo 0x1825231fc <+68>: add x24, x0, #0x20 // x24 == os_unfair_recursive_lock_s 0x182523200 <+72>: mov x0, x24 0x182523204 <+76>: mov w1, #0x0 0x182523208 <+80>: bl 0x188096910 // os_unfair_recursive_lock_lock_with_options crash 调用栈
调用 object_getClass 获取 Class,调用 object_getIndexedIvars 获取到 _notifyingInfo,_notifyingInfo + 偏移量 0x20 获取 os_unfair_recursive_lock_s,崩溃的原因是这把锁的 owner 损坏了,lock 也是一个结构体,ower 也是根据 offset 获取的。
结论
从崩溃的上下文来看,最可能出问题的是获取 _notifyingInfo,因为只有 KVO Class 才能获取到 _notifyingInfo 这个结构体,如果在调用 __NSSetBoolValueAndNotify 的过程中,在其它线程监听被移除,此时 object_getClass 取到的不是 KVO Class 那后续再根据 offset 去取 lock,这个时候就有可能发生上述崩溃。
线下暴力复现验证了上述猜测。
- (void)start { __block KVObject *obj = [KVObject new]; dispatch_async(dispatch_get_global_queue(0, 0x0), ^{ for (int i = 0; i < 100000; i++) { [obj addObserver:self forKeyPath:@"value" options:0x7 context:nil]; [obj removeObserver:self forKeyPath:@"value"]; } }); dispatch_async(dispatch_get_global_queue(0, 0x0), ^{ for (int i = 0; i < 100000; i++) { obj.value = YES; obj.value = NO; } }); } - (void)observeValueForKeyPath:(NSString *)keyPath ofObject:(id)object change:(NSDictionary<NSKeyValueChangeKey,id> *)change context:(void *)context {}
解决这个问题的思路就是保证线程安全,我们在线上断点找到了 removeObserver 的代码,将 removeObserver 和触发监听的代码放在了同一个串行队列。当然如果 removeObserver 在 dealloc 里面,理论上也不会出现这类问题。
__NSSetxxxValueAndNotify 系列方法都有可能会触发这个崩溃,类似的问题可以按照相同的思路解决。
00000000004e05cd t __NSSetBoolValueAndNotify 00000000004e0707 t __NSSetCharValueAndNotify 00000000004e097b t __NSSetDoubleValueAndNotify 00000000004e0abc t __NSSetFloatValueAndNotify 00000000004e0bfd t __NSSetIntValueAndNotify 00000000004e10e7 t __NSSetLongLongValueAndNotify 00000000004e0e6f t __NSSetLongValueAndNotify 00000000004e0491 t __NSSetObjectValueAndNotify 00000000004e15d5 t __NSSetPointValueAndNotify 00000000004e1734 t __NSSetRangeValueAndNotify 00000000004e188a t __NSSetRectValueAndNotify 00000000004e135f t __NSSetShortValueAndNotify 00000000004e19e8 t __NSSetSizeValueAndNotify 00000000004e0841 t __NSSetUnsignedCharValueAndNotify 00000000004e0d36 t __NSSetUnsignedIntValueAndNotify 00000000004e1223 t __NSSetUnsignedLongLongValueAndNotify 00000000004e0fab t __NSSetUnsignedLongValueAndNotify 00000000004e149a t __NSSetUnsignedShortValueAndNotify 00000000004de834 t __NSSetValueAndNotifyForKeyInIvar
以上就是99% iOS开发都不知道的KVO崩溃分析详解的详细内容,更多关于iOS开发KVO崩溃的资料请关注脚本之家其它相关文章!
相关文章
详解iOS开发 - 用AFNetworking实现https单向验证,双向验证
这篇文章主要介绍了详解iOS开发 - 用AFNetworking实现https单向验证,双向验证,具有一定的参考价值,感兴趣的小伙伴们可以参考一下。2016-12-12举例讲解Objective-C中@property属性的用法
这篇文章主要介绍了Objective-C中@property属性的用法,包括@property的属性关键字的整理,需要的朋友可以参考下2016-03-03
最新评论